# AI Economy Part 1: The Physics of AI

> Machine-readable edition for AI agents. Canonical: https://maxwel.xyz/en/research/reports/2026-07-ai-economy/part-1 (PDF, podcast and video available there).

| | |
|---|---|
| **Series** | AI Economy, research reports as the foundation for a presentation |
| **Report** | 1 · *Understanding (foundations):* what AI physically depends on, the causal chain from unbounded demand through chips, raw materials and electricity to the power wall, and the solution in orbit |
| **Status** | 08/26/2026 |

> **Note:** This report by **Maxwel Consulting** was created with the AI-powered research pipeline **Maxwel Research** (Claude Fable 5) under human supervision and curation. Despite careful work, individual figures may be inaccurate, outdated or incomplete; Maxwel Consulting accepts no liability for accuracy, completeness or currency. This report does not constitute investment, legal or tax advice. The field moves at weekly pace; vendor and single-source figures are labeled as such. Factual errors: support@maxwel.xyz.

## Key findings

1. **Demand for AI is effectively unbounded.** Token consumption, meaning the number of word building blocks the models process, grew roughly 330-fold in two years, every price cut *increases* usage (Jevons paradox), and the capital appetite is driving the industry to the stock exchange as a bloc in 2026.
2. **The chip side is delivering, and what is really scarce is only memory (HBM, the fast memory sitting right next to the chip), packaging (the joining of chip and memory) and fabrication slots.** The chain is a sequence of quasi-monopolies (ASML/Zeiss/Trumpf → TSMC → SK Hynix → Nvidia), and three of the top positions, the EUV optics, the EUV laser and semiconductor polysilicon, sit with European suppliers (Zeiss, Trumpf, Wacker).
3. **China's counter-program has been slowed, not stopped, because memory caps the build-out.** The compute balance is that the US is chip-rich and power-limited, China is power-rich and chip-limited, and Europe (4.8 %) is neither.
4. **Electricity is the binding constraint (AI power wall), and inference is taking over.** Running the models needs \~93 GW by 2030 versus \~62 GW for training, round the clock, close to the user and with daily peaks; one gigawatt is roughly the output of one large power plant.
5. **The structural solution lies in orbit.** AI satellites in sun-synchronous orbits have effectively unlimited energy and cool via infrared radiators; commercial scale arrives in the 2030s.

### The series at a glance

| Part | Title | Guiding question |
|---|---|---|
| **1 (this one)** | The physics of AI | What does AI physically depend on, and where is the limit? |
| 2 | The models of AI | How do models come into being, who builds them, how do you work with them? |
| 3 | The application of AI | Where does AI concretely help the business, and how do you own it yourself? |

---

The whole world is betting on artificial intelligence right now. Companies are building data centers, states are subsidizing chips, and the four big tech companies together are investing about 725 billion dollars in 2026.[^1] For anyone using AI in a business, something concrete hangs on this, namely the prices of the models, their availability at peak times, and the question of whose chips, grids and raw materials their own usage depends on. The first part of the three-part AI Economy series covers the physical chain behind it, because AI is first and foremost an infrastructure problem. The report follows this chain section by section, from effectively unbounded demand through chips as the first answer, the raw materials from which everything is built, and the electricity demand to the power wall, meaning the limit at which it is no longer the chips but the available electricity that constrains the build-out, on to the energy deals as a bridge, and at the end to the possible solution, the **AI satellites**, which dodge the power wall by going into orbit.

## Demand: effectively unbounded

**Consumption is growing by whole orders of magnitude.** In May 2024, Google processed roughly 9.7 trillion tokens per month, meaning the word building blocks into which the models split every text. In May 2025 it was \~480 trillion, and in May 2026 **over 3,200 trillion**, a \~330-fold increase in two years and 7× in the last year alone (a self-reported figure, but consistent across three keynotes).[^2] The neutral router OpenRouter, which forwards requests to many model providers, grew \~15× over the same year,[^3] and Google's model APIs now process \~22 bn tokens **per minute** (+38 % in one quarter).[^4] The evidence of this demand keeps piling up at every level. ChatGPT crossed **1 bn weekly active users** (official since August 2026, and the billion now applies per week rather than per month),[^5] Anthropic grew at **80× annualized** in the first quarter of 2026 ("we had planned for 10×"),[^6] Nvidia's data-center revenue rose to 75 bn $ per quarter (+92 %, audited),[^7] and the four large US corporations together are investing **\~725 bn $** in AI infrastructure in 2026 (+77 %).[^1] A European provider does not appear in any of these statistics.

**The capital appetite behind this is now bursting the private markets.** After the largest venture rounds of all time (OpenAI with 122 bn $ at an 852 bn $ valuation,[^8] Anthropic with 65 bn $ at **965 bn $**, since May 2026 the world's most valuable AI company),[^9] the industry is pushing toward the **stock exchange** in 2026. Anthropic filed its IPO confidentially with the US Securities and Exchange Commission in June;[^10] OpenAI, after its own filing,[^11] is now leaning toward a debut only in 2027 and instead let employees sell \~7 bn $ of shares in August (valuation unchanged at 852 bn $).[^12] Chip builder Cerebras has been listed on Nasdaq since May (proceeds 5.55 bn $, +68 % on day one),[^13] and CoreWeave had already led the way in 2025 as the first major AI-infrastructure IPO (trading at over 2× its issue price).[^14] China shows the same picture in the same year. Zhipu and MiniMax have been listed in Hong Kong since January[^15] (Zhipu up more than 10× since debut),[^16] memory manufacturer CXMT completed the **largest China chip IPO since SMIC** in July (proceeds doubled to \~8.6 bn $, debut +466 %, at times the most valuable China-listed company),[^17][^18] and robotics maker Unitree followed in August (+460 % on day one).[^19][^20] 2026 is on track to become the largest IPO year of all time, driven by AI;[^21][^22] in venture capital, AI recently accounted for \~87 % of US volume.[^23] Growth on this scale simply can no longer be financed privately.

**That this demand does not saturate is explained by the Jevons paradox.** As with coal in the 19th century, efficiency does not reduce consumption but fuels it. Satya Nadella wrote after the DeepSeek shock, "Jevons paradox strikes again! … we will see its use skyrocket, turning it into a commodity we just can't get enough of."[^24] The numbers bear him out, because the price per "unit of intelligence" is falling **9- to 900-fold per year** depending on the yardstick (GPT-4 level got over 40× cheaper in two years),[^25] and total token revenue is still growing. The hardest piece of evidence came from DeepSeek in April 2026. When the provider cut prices by up to 90 %, token volume rose **by +297 % on the first day**, and demand became so large that DeepSeek **had to introduce peak-hour surcharges**.[^26] The price cut therefore created a capacity bottleneck, and in August 2026 came the reversal, because DeepSeek announced **price increases** for the V4 generation, and demand is carrying them.[^27]

**Every new mode of use multiplies consumption further.** Thinking models, which produce a reasoning text before answering, need roughly eight times as many tokens per query as simple ones;[^28] **agents, which work through a task autonomously in many steps, consume up to \~1,000× more** per task than a chat, because the context is re-sent at every step;[^29] a 10-second AI video costs \~1.30 $ of raw compute,[^30] and Sora reportedly burned through \~15 mn $ per day until OpenAI shut down the app to redirect the GPUs, meaning the compute chips, onto more profitable products.[^30] Even the world's richest provider therefore has to **ration** demand.

**The ceiling lies at the entire cognitive work of humanity.** McKinsey estimates the annual value potential of generative AI at 2.6–4.4 tn $ (model estimate),[^31] and practically every lab and corporate chief publicly declares themselves **"compute-constrained"**, meaning limited by computing power. Altman speaks of being "out of GPUs" and "our GPUs are melting" and names 100 mn GPUs as the long-term target,[^32] Amodei says "80× is too hard to handle",[^6] Microsoft's CFO calls the company capacity-constrained "at least through 2026", despite \~190 bn $ in investment,[^33] Pichai speaks of being "compute constrained in the near term",[^33] and Huang says "demand has gone parabolic", which is the PR of a beneficiary but consistent with everything else.[^34]

> **The thesis for everything that follows is that demand for intelligence is not limited by need but only by supply.** And supply no longer hangs on the chip but on the electricity grid. The following sections make the case step by step.

## Chips: the first answer to demand

The industry's first answer to this demand is silicon, and to understand where things get stuck, it is worth a look at the fundamentals.

## Why GPUs, and why memory is the bottleneck

A processor (CPU) has a few dozen general-purpose compute cores; a **GPU**, originally a graphics chip, has tens of thousands of simple cores that execute the same arithmetic operation in parallel on huge amounts of data, and that is exactly the pattern of the matrix multiplications that make up neural networks. But the actual bottleneck is not the computing but the **memory**. When answering, all the model weights have to be read out of memory for each generated word building block, and the compute cores therefore mostly wait for data ("memory wall"). The answer to this is called **HBM** (High Bandwidth Memory), a fast memory sitting right next to the chip, in which 8–16 layers of memory chips are stacked vertically beside the compute unit with an extremely wide interface. The capacity series shows how the arms race is playing out.[^35]

| GPU | Year | Memory | Bandwidth |
|---|---|---|---|
| Nvidia A100 | 2020 | 80 GB HBM2e | \~2.0 TB/s |
| H100 | 2022 | 80 GB HBM3 | 3.35 TB/s |
| H200 | 2024 | 141 GB HBM3e | 4.8 TB/s |
| B200 (Blackwell) | 2024/25 | 192 GB HBM3e | 8 TB/s |
| B300 (Blackwell Ultra) | 2025 | 288 GB HBM3e | 8 TB/s |
| **Vera Rubin** | H2 2026 | **288 GB HBM4** | **22 TB/s** (manufacturer figure) |

HBM4 doubles the interface width (2,048 bits per stack, \~2 TB/s, 36 GB per stack in series).[^36][^37] Three manufacturers from two countries share the market, **SK Hynix (South Korea) with \~58 %, Samsung and Micron with \~21 % each** (Q1 2026),[^38] and Samsung has caught back up with Micron and was the first to deliver HBM4 for Nvidia's Rubin.[^39] Europe does not appear in it, because since the insolvency of the Dresden DRAM manufacturer **Qimonda** (2009) there has been no European memory production.[^40] Alongside Nvidia, others build their own accelerators, Google the **TPU v7 "Ironwood"** (the first inference-optimized TPU, meaning designed for running operations rather than for training)[^41][^42] and AMD the MI450 as the first tranche of a 6 GW OpenAI deal.[^43]

## The scaling ladder: from chip to gigawatt campus

AI compute scales in tiers, and each tier multiplies the electricity demand. For orientation, a gigawatt (GW) is a measure of power, and 1 GW is roughly the output of one large power plant; the tiers look like this:

- **The first tier is the chip.** One Rubin GPU has 336 bn transistors, \~50 PFLOPS of compute and **1,800–2,300 W** of power draw and can only be cooled by liquid.[^44]
- **The second tier is the server.** Classically 8 GPUs sit in one server (8× H100 = 640 GB of memory).
- **The third tier is the rack.** A Vera Rubin NVL72 connects 72 GPUs and 36 CPUs with **20.7 TB HBM4** and 3.6 EFLOPS; the Blackwell predecessor draws \~120 kW *per cabinet*.[^44]
- **The fourth tier is the campus.** xAI **Colossus 2** (Memphis) is regarded as the **first gigawatt data center**, with \~550,000 Blackwell GPUs under construction and on track for \~1 mn GPUs / 2 GW;[^45][^46] OpenAI's **Stargate Abilene** is targeting 1.2 GW with >450,000 GPUs,[^47] the Stargate program overall \~10 GW by 2029.[^48]
- **Europe's scale is small by comparison.** The largest European AI computer **JUPITER** (Jülich, Europe's first exascale machine, \~24,000 Nvidia superchips)[^49] would be roughly **20× smaller** than Colossus 2, and the planned EU "AI Gigafactories" (\~100,000 GPUs per site) have yet to break ground.[^50][^51][^52]

## Who builds all of this: the vulnerable manufacturing chain

Behind every GPU stands a highly concentrated chain. **Nvidia (US) designs, TSMC (Taiwan) fabricates**, with \~72 % market share in contract manufacturing overall and **>90 % on the most advanced processes**; practically every AI accelerator in the world (including those from AMD, Google, Amazon) runs through TSMC.[^53] The tools for it are supplied by a single company, and this is Europe's strongest card. **ASML** (Netherlands) holds the world monopoly on EUV lithography machines. **Lithography** is the name of this step, in which the circuit structures are photographically exposed onto the silicon wafer, and EUV stands for extremely short-wavelength light, which allows the finest structures. The machine is a **European co-production**. The centrepiece, the complete **mirror optics**, comes exclusively from **Zeiss SMT** (Oberkochen), and these mirrors are considered the most precise manufactured objects in the world, because if such a mirror were the size of Germany, its highest unevenness would be \~0.1 mm (manufacturer figure); ASML has taken a 24.9 % stake in Zeiss SMT to secure itself against this dependency.[^54] The **light source** is built exclusively by **Trumpf** (Ditzingen), a high-power CO2 laser that ignites 50,000 tin droplets per second into plasma.[^55] And ASML manufactures in Germany itself, because **ASML Berlin** (the former Berliner Glas, acquired in 2020, >1,700 employees) builds wafer stages, reticle chucks and mirror blocks,[^56] and **Jenoptik** (Jena) supplies further optical components. **Without these suppliers, therefore, no single EUV machine exists**, and on the link in the chain that China demonstrably cannot copy, two of the three monopolies sit with Zeiss and Trumpf. The newest High-NA class costs **\~350–400 mn $ per machine**, and fewer than twelve exist worldwide.[^57]

The scales of manufacturing are correspondingly large. TSMC's large plants are officially called **"Gigafab"** (≥100,000 wafers/month), and the **Gigafab cluster** under construction **in Arizona** costs **165 bn $** (six fabs + two packaging plants).[^58] A single modern fab costs 20–30 bn $ and takes **3–5 years from groundbreaking to serial volume**. Europe's manufacturing reality looks modest alongside this. TSMC is building with Bosch, Infineon and NXP in Dresden (**ESMC**, \~10 bn €), but for 28–12 nm industrial and automotive chips and **not for AI accelerators**;[^59] Intel's Magdeburg project is buried,[^60] and the EU Chips Act target (20 % of the world market by 2030) is regarded as unattainable.[^61] To be distinguished from this is **"Terafab" [announcement]**, Elon Musk's chip project announced in March 2026 (Tesla/SpaceX/xAI, \~25 bn $; Intel involved as foundry partner, meaning contract manufacturer, since April)[^62][^63] with the goal of "1 terawatt of AI compute per year", which would be **\~100× the current global production** of AI chips (in real terms \~5–10 GW of connected capacity per year, rough estimate). In August 2026 the first construction phase was fixed, Grimes County in Texas with a 16.8 bn $ initial investment, and the announcement has thereby become a financed project that shows how seriously the industry is taking unbounded demand.[^64]

## The bottlenecks in 2026: memory, packaging, fab slots

On the chip side, three things are **actually** scarce in 2026:
1. **HBM memory is missing in the first place.** Micron is sold out for the calendar year, the HBM4 serial ramp is only just beginning,[^65] and Nvidia's Rubin output in 2026 may as a result be capped at \~200–300 k GPUs (manufacturer figure).[^66][^67]
2. **CoWoS packaging remains tight through year-end.** CoWoS is the joining of GPU and HBM on a silicon carrier at TSMC, and a gap of \~10 % remains there at the end of 2026 (TrendForce).[^68]
3. **Fab slots at the leading edge are scarce in general.** TSMC's chief expects the tightness to last through 2027.[^69]

The raw-material side of chips, by contrast, is **manageable but vulnerable** in 2026. The neon crisis (Ukraine supplied \~90 % of US demand) was averted through stockpiles and new sources;[^70] high-purity quartz for silicon crucibles comes \~70–90 % from a single region (Spruce Pine, North Carolina, where a hurricane took the mines offline for weeks in 2024);[^71] gallium and germanium are under Chinese licensing (details in the raw materials section below, likewise the water footprint of the fabs).

## China's counter-program, and where Europe stands in it

Under US export controls, China is working on complete self-sufficiency across the chip chain, with the **Big Fund III** (\~47.5 bn $, larger than both predecessors combined)[^72] and hard targets. Europe's counterpart, the **EU Chips Act**, mobilizes \~43 bn €, a similar order of magnitude on paper, but spread across years and member states and without a single leading-edge project, meaning without a plant for the most advanced processes;[^61] the **"Chips Act 2.0"** presented in June 2026 for the first time explicitly targets a leading-edge fab for AI chips, but for now is a Commission proposal.[^73] The honest state of affairs in mid-2026 looks like this, link by link:

- **Logic manufacturing really stands at \~7 nm.** SMIC produces the 7 nm class in serial volume (Huawei Ascend 910C, Kirin phone chips), via DUV multi-patterning with the older laser technology and entirely without EUV. Chinese "5 nm serial production" reports are embellished, because independent chip analyses measure something closer to the TSMC-6 nm class, with yields estimated at 20–60 % (TSMC >90 %) and wafer costs 40–50 % higher.[^74]
- **The toughest gap of all is lithography.** A secretive Shenzhen lab has a reverse-engineered **EUV prototype** (Reuters), which produces EUV light but **has not yet exposed a single chip**; serial readiness is expected in \~2030 at the earliest.[^75][^76] One rung below, serial hardware has existed since summer 2026 for the first time, because Shanghai is producing its own immersion DUV machines in small series (\~5 units in 2026, \~20 planned for 2027), technically \~four generations behind ASML, symbolically relevant, quantitatively not yet (China imported 95 DUV systems in 2025 alone).[^77] ASML's chief puts the EUV gap at **10–15 years**.[^78] It is the one link in which **Europe** is the unassailable world market leader, because ASML, Zeiss and Trumpf *are* this gap.
- **Memory is the actual cap on the build-out.** CXMT masters HBM2 but is missing its 2026 HBM3 target (yield problems, \~3–4 years behind SK Hynix, which is meanwhile ramping HBM4).[^79] Huawei is living off **\~13 mn hoarded HBM stacks** from abroad (enough for \~1.6 mn Ascend accelerators), and domestic production in 2026 contributes only \~250–300 k.[^80] **It is not wafers but memory that caps China's AI chip build-out**,[^81] and it is compensated for by cluster scaling (Atlas 950 SuperCluster with 520,000+ chips).[^82]
- **In tools there are strengths and holes.** Naura has risen to become the world's fifth-largest equipment maker, and AMEC etchers run right into 5 nm lines (also at TSMC). The localization share in China's fabs is \~35 %, and the target is 70 % by 2027.[^83] The holes are lithography, metrology and inspection (\~10–25 % depending on source) and ion implantation.
- **Politics substitutes for the missing technology.** The old 70 % self-sufficiency target (2025) was **clearly missed**[^84] (in reality, by analyst estimates, roughly 20 to 30 percent, depending on how one counts).[^85] The answer to this is a forced home market. Since November 2025, foreign AI chips have been **banned** in state-funded data centers,[^86] and even the H200, released again by the US, is blocked by Chinese customs, so that Nvidia's China data-center revenue is effectively zero.[^87]

**The compute balance shows a factor of \~5 in installed base, but the mirror image in electricity.** Epoch AI attributes **\~74.5 %** of the world's recorded AI supercomputer compute to the US, **\~14 %** to China, roughly one-fifth of the US.[^88] (An honest caveat is that the dataset captures only 10–20 % of global capacity and barely any of China's gray imports, so China's share is more a lower bound.)[^89] In **build-out**, the gap is widening, because Nvidia shipped \~5–6 mn accelerators in 2025 (analyst forecast for 2026 \~7.5 mn),[^90][^91] and Huawei reaches \~4 % of Nvidia's compute output.[^92] It is exactly the opposite for **electricity**. At end-2025 China had **3.89 TW** of installed generation capacity, meaning 3,890 GW (+16 % in one year; build-out \~540 GW, with more solar added than the rest of the world combined),[^93][^94] the US \~1.25–1.3 TW, and it is fighting for every gigawatt (see the power wall below).[^95] **The US is chip-rich and power-limited, China is power-rich and chip-limited.** Europe is barely visible in these statistics, because **the EU-27 comes to a combined 4.8 %** at Epoch,[^88] even though it would have \~1.17 TW of generation capacity (end-2024, Eurostat).[^96] So Europe is not primarily short of electricity, but of chips, campuses and capital decisions.

**How China nevertheless delivers frontier models without its own chip chain is explained by four mechanisms:**
1. **Training needs little compute, operation a lot.** A frontier training run, meaning the training of a top-tier model, gets by with tens of thousands of GPUs, and China has those (hoarded H800/H100/H20). DeepSeek V3 was, on the company's own account, trained on 2,048 H800s, and the final run cost \~5.6 mn $ of compute[^97] (SemiAnalysis estimates the lab's overall access at \~50,000 Hopper GPUs).[^98] The millions of GPUs the US installs, by contrast, sit predominantly in **operation** for billions of users.
2. **Open weights offload the compute to the world.** DeepSeek and GLM publish the model weights, and the compute is then done worldwide on Nvidia hardware, in US clouds and on company machines. The rest of the world thus bears the operating load. That things nevertheless get tight domestically is shown by DeepSeek's peak-hour surcharges (see the demand chapter).
3. **Efficiency replaces part of the raw power.** Mixture-of-Experts activates only a fraction of the model per word building block (V3 37 of 671 bn parameters, V4 Flash 13 of 284 bn),[^97][^99] on top of which come FP8 training with lower numerical precision and memory-frugal attention methods, and scarcity has demonstrably forced invention here.
4. **Breadth replaces the peak, supplemented by the gray zone.** Weaker chips are wired into larger clusters (Atlas 950; possible because electricity is not scarce in China). Alongside this comes well-documented gray procurement, because according to FT reporting, \~1 bn $ of B200/H100/H200 was smuggled into China in just three months of 2025,[^100] and the Megaspeed case (\~2 bn $ in Nvidia purchases, cloud access for Chinese customers via Malaysia/Indonesia)[^101] and several US and Singapore indictments corroborate the pattern.[^102]

**The reality check for the Ascend comes out mixed.** In 2025, DeepSeek failed in its attempt to **train** R2 on Huawei Ascend (unstable, slow interconnects, immature software); training went back onto Nvidia, Ascend remained inference (journalistic reporting, anonymous sources).[^103] In 2026, Zhipu by contrast reports having trained GLM-5 and GLM-5.2 **entirely on Ascend** (own account, not independently verifiable),[^104] and if that stands up, it would be the first evidence that the Huawei chain carries frontier training. **The phrase "Chinese chip" calls for caution, though.** The Ascend is Chinese-*designed* (Huawei/HiSilicon), but not entirely Chinese-*built*, because teardowns found **TSMC-fabricated dies**, meaning the actual chip pieces, still in Ascend 910C samples (\~2.9 mn compute dies, procured around 2023/24 via the intermediary Sophgo, analyst estimate; plus older CPU dies from \~2020),[^105][^106] and the HBM memory on them comes from Samsung and SK Hynix (the hoarded \~13 mn stacks from above). "Trained on Ascend" therefore means **Nvidia-independent, but not TSMC- and memory-independent**.

**In sum, the export controls have not stopped China, but have measurably slowed it and made it more expensive**; on aggregated AI compute the gap even widens on the CFR calculation (Huawei \~4 % of Nvidia's compute output in 2026, trend falling).[^92] China's most effective counter-weapon is not technology but the **raw-material lever** (next section) and the ordered domestic market. For Europe, both of these are instructive. The chip chain is a sequence of quasi-monopolies (ASML → TSMC → SK Hynix → Nvidia), and **Europe sits irreplaceably in exactly one link** (lithography), and in none of manufacturing, memory or chip design. That one link gives Europe bargaining power; everywhere else it remains dependent. And precisely this link became an open point of contention between the blocs in August 2026. With the **MATCH Act**, the US is preparing to force the Netherlands into a near-total ban on ASML's China business, not just on new DUV sales but also on the **servicing of machines already delivered**, with the threat of sanctions for non-compliance.[^107][^108] The Hague is publicly pushing back ("export control works best out of conviction, not by decree from across the border"), and the prospects of success are considered limited;[^109] China's share of ASML's system revenue has already fallen from 33 % (2025) to 14 % (Q2 2026).[^108] Europe's monopoly is thus no longer just bargaining power, but has itself become an object of the chip conflict.

## Raw materials: what the chain is built from

The raw-materials question has two levels, namely what sits **in the chip itself** and what the **campus around it** devours.

## In the chip: more than 60 elements

A chip from the 1980s made do with **12 elements**; a modern one uses **more than 60**, more than half of the periodic table (NRC time series on Intel data).[^110] The most important follow here with an honest risk placement:

- **Silicon is the base, with a European trump card.** From quartz, raw silicon is made, and from that **semiconductor polysilicon** of ≥ 99.999999999 % purity ("11 nines"). This top-tier market is tiny (\~2 % of the polysilicon volume; the rest is solar grade, dominated by China with \~95 %) and is controlled by a duo, because **Wacker (Germany, world market leader in this segment) and Hemlock (US, No. 2) together supply \~¾ of the world's semiconductor polysilicon**;[^111] in bulk polysilicon for solar, by contrast, Wacker is only ranked 8th behind China's giants, so Europe holds not the volume but the purity peak.[^112] Alongside chip exposure (the lithography with the ASML/Zeiss/Trumpf machines, see chips section), this is Europe's second irreplaceable link in the chain.
- **On the inside sit several metals with risks of their own.** **Copper** interconnects (since IBM 1997, instead of aluminum),[^113] **tungsten** for the contacts (China \~80 % of extraction, export controls since February 2025),[^114] **cobalt** in the most advanced interconnects (DR Congo \~76 %),[^115] **tantalum** as a diffusion barrier (a conflict raw material) and **hafnium** as an atomic-layer-thin gate oxide (since 2007; only a few dozen tonnes a year worldwide arise as a by-product of zirconium refining, uncritical in volume, structurally inelastic).[^116]
- **Dopants** (boron, phosphorus, arsenic, antimony) are what first turn silicon into transistors, in tiny amounts but with no free pass, because **antimony** has been under Chinese export controls since August 2024;[^117] its price at times nearly tripled, but since the controls pause it has given back a good half of the rise (a price, not a volume, risk).[^118]
- **Rare earths sit in the fabrication process itself.** **Cerium oxide** polishes the wafers mirror-smooth (China refines \~90 % of rare earths),[^115] and **lanthanum** sits in the gate stack.
- **Packaging hangs on solder and thus on the tin market.** **Tin** holds the world's electronics together as solder (\~half of the world market),[^119] a small market with concentration risk, because Myanmar's mining stop, Indonesian export restrictions and fund speculation drove the price to record levels in 2026, with a nominal all-time high in January (\~53,500 $/t)[^120] and \~56,000 $/t in August.[^121] On top of that come gold, silver and copper bonding wires.

## Process consumables: the quiet chokepoints

Fabricating chips continually consumes specialist substances whose origin is highly concentrated. **Neon** for the exposure lasers (Ukraine supplied \~50 % of world demand, up to \~90 % of US semiconductor demand),[^122] **helium** as a coolant and carrier gas (US \~43 %, Qatar \~33 %; the Hormuz crisis in 2026 abruptly cut off \~30 % of world supply),[^123] **high-purity hydrofluoric acid and photoresists** almost exclusively from Japan (\~90 % for EUV photoresists; the Japan-South Korea dispute in 2019 proved that even *allies* can turn off this tap),[^124] **gallium** for power electronics (China \~98–99 % of raw production)[^125] and **germanium** for high frequency and photonics (China \~60 %).[^126]

## The campus around it: copper, magnets, water

**Copper** is the systemic risk at this level. AI campuses require 3–4× the amount of copper (power + liquid cooling);[^127] BNEF projects a deficit of \~6 Mt by 2035,[^128] the IEA now even warns of a 30 % gap,[^129] and the price set an all-time high in August 2026 (\~15,000 $/t).[^130] **Water** is the overlooked resource alongside, more precisely a *local* bottleneck. A large chip fab draws in **20–38 mn liters** per day (plus 30–50 MW of electricity), usually from the drinking-water grid, and refines most of it into **ultrapure water** for rinsing the wafers, many orders of magnitude purer than drinking water.[^131] The polluted process wastewater is treated on site and increasingly circulated (modern fabs recycle 85–90 %, manufacturer figures); water is really "consumed" mainly through **evaporation in cooling towers**, which is also true for data centers. The core problem is therefore not destruction but **competition for local drinking water in dry regions** (Taiwan let fields go dry during the 2021 drought in order to supply the fabs);[^132] treated wastewater as a substitute source is only gradually being tapped. **Rare earths** sit in the magnets of the cooling pumps, and China controls \~85–90 % of processing and \~90 % of magnet production.[^133]

## The geopolitical lever

The Chinese export controls on gallium, germanium and rare earths are currently **suspended, not lifted**. More precisely, only the tightened October-2025 package is on hold (until November 27, 2026), and the underlying regime of licensing requirements (Ga/Ge since 2023, rare earths since April 2025) remains active.[^134][^135] It is a "pause for recalibration, not reconciliation",[^136] and at the same time China's most effective lever in the chip conflict (see above). **Europe is the most exposed bloc at this lever.** According to the European Commission, 98 % of the rare earths Europe needs come from a single supplier, namely China.[^137] The **Critical Raw Materials Act** (in force since May 2024) sets targets for 2030 (among them 10 % own extraction, 40 % own processing, max. 65 % per third country),[^138] and initial projects are running, because Solvay (Belgium) has been producing magnet rare earths in French La Rochelle again since April 2025, aiming for 30 % of EU demand by 2030;[^139] against China's scale that is a beginning for now, no more.

## Compute growth, and what the chips demand in electricity

Along with the chips, compute continues to grow exponentially. The training compute of frontier models, meaning the amount of computation for training the top-tier models, has been doubling roughly **every 5–6 months** since 2020;[^140] the cost of a top-end training run is growing at \~2.4×/year, and Epoch projects **>1 bn $ per run by 2027**.[^141] (That is not the electricity bill, because the lion's share is amortized hardware, meaning accelerators, servers and cluster networking; electricity typically accounts for only single-digit percent. Electricity is the *availability* limit, not the cost driver.) Physically, scaling could be pushed on to \~2·10²⁹ FLOP by \~2030, meaning arithmetic operations (≈ 5,000× GPT-4).[^142] **But all these chips want to be supplied**, and a single training campus of this magnitude demands **1–5 gigawatts** of electrical power, meaning the output of one to five large power plants.[^142]

**And inference, the running operation, is growing even more steeply.** Token volume is currently doubling roughly **every 4 months** (Google's \~330× in two years, see demand chapter[^2]), faster than training compute with its 5–6 months. Countervailing this, consumption **per query** is falling drastically, because Google puts the median text prompt at **0.24 Wh** (≈ 9 seconds of television; 0.26 ml of water), **33× less energy than a year earlier** (own figure, methodology disclosed).[^143][^144][^145] Net, growth still remains exponential, because volume and complexity more than offset efficiency (Jevons, see demand chapter). McKinsey projects the **inference power requirement to go from \~21 GW (2025) to \~93 GW (2030)**, growth of \~35 %/year and markedly faster than training (\~62 GW, \~22 %/year); from \~2026/27 onwards, inference is permanently the larger block.[^146] The structural difference lies in the load profile, because training is episodic and movable, whereas inference runs **round the clock, latency-bound, meaning dependent on short response times and therefore close to the user, with daily peaks** (DeepSeek's peak-hour surcharges are the first price signal[^26]). **Training built the large campuses, and from now on it is inference that pays the electricity bill.**[^147]

## AI power wall: electricity is the binding limit

Several independent analyses in 2026 arrive at the same conclusion, "The binding link is now the power grid, not silicon."[^148] Epoch AI ranks the four scaling limits (electricity, chip manufacturing, data, latency) and sees **electricity as the first binding limit, then chip production**.[^142] Concretely, according to the IEA, global data centers consumed \~485 TWh in 2025 (\~1.7 % of world electricity; terawatt-hours measure consumption over the year, gigawatts measure power at a given moment) and are set to roughly double to \~950 TWh (\~3 %) by 2030, growing four times faster than any other sector.[^149][^150] Anthropic's own estimate is that the US AI sector will need **≥ 50 GW** of additional capacity by 2028 (about 50 large reactors).[^151]

The bottleneck is physical, not financial. \~11 GW of announced US capacity is stuck at grid connection, **transformers and switchgear have multi-year lead times** (details below), and interconnection queues run over 5 years.[^152] Europe is facing the same wall on a smaller scale. In the greater **Dublin** area, Europe's densest data-center location, data centers already account for **22 % of Irish electricity** (2024, official),[^153] and a **connection freeze on new data centers** had been in place there since 2022;[^154] since 2026 it has been lifted, but against hard conditions, because new data centers must be able to generate or store their full load themselves, source 80 % from new Irish renewables within six years and feed back into the grid on demand, and the Dublin grid itself remains extremely tight.[^155] Frankfurt and Amsterdam are also throttling via grid and land-use conditions.

## Energy: the deals, the bottlenecks, the workarounds

Gas covers the next few years, nuclear is still paper for 2027 and beyond, and gas turbines step in short-term. **Crusoe** (US), the data-center developer behind the Stargate Abilene campus, originally launched with data centers on flared oilfield gas, ordered 29 GE Vernova aeroderivative turbines (\~1 GW), meaning turbines derived from aircraft engines,[^156] and Abilene itself is building \~1 GW of own generation.[^157] **But the turbines themselves have become the bottleneck**, and the market is a three-country oligopoly of **GE Vernova (US)**, **Siemens Energy (Germany)** and **Mitsubishi Power (Japan)**. GE Vernova, with a **116 GW turbine backlog**, is by now **sold out through 2031**;[^158][^159] Siemens Energy, one of the boom's big beneficiaries with gas-turbine manufacturing including in Berlin, is reporting record order backlogs across the board[^160] (gas-turbine backlog \~69 GW in August 2026,[^161] plus a record 51 bn € in the grid business alone)[^162] and cites over 3 years' waiting time for new orders; Mitsubishi is also at capacity limits.[^163] Crusoe's deal therefore secured above all scarce fabrication slots, because even the quickly *installable* aeroderivative machines now have 3–5 years of ordering lead time.[^164]

In parallel, several large nuclear deals are running. AWS-Talen secures up to 1.92 GW direct from the 2.5 GW **Susquehanna** nuclear plant in Pennsylvania (Amazon bought the adjoining data-center campus for it), Microsoft-Constellation the Three Mile Island restart (\~835 MW, target 2027, brought forward)[^165] and Google-Kairos the first US corporate SMR fleet, meaning small modular reactors.[^166] But the SMR market is still small (\~6.9 bn $ in 2025), and the nuclear answer is coming too late for acute demand.[^167] There is no European counterpart to these deals; the closest is **France**, which is aggressively marketing its nuclear surplus to AI data centers (Macron at the Paris AI Summit with "plug, baby, plug").[^168]

**Globally, solar has long been the build-out engine.** China's record additions in 2025 (\~540 GW) were dominated by solar (378 GW, 58 % of world additions),[^93][^94] and the largest US additions since 2002 (53 GW) likewise consisted predominantly of solar plus battery storage.[^95] For data centers, solar is the cheapest electricity, but not always-on electricity, because without large storage an installation delivers only about a quarter of the year's hours. That is why the hyperscalers, meaning the large cloud companies, have so far bought solar mostly on a book basis via the grid, while the 24/7 baseload hangs on gas and nuclear.

**The battery-buffered solar data center is long since not a theory.** In Nevada, a Crusoe AI data center has been running since June 2025 on a microgrid of **12 MW of solar and 63 MWh of decommissioned EV batteries** (Redwood, founded by Tesla co-founder JB Straubel, the world's largest second-life battery deployment),[^169] with **99.2 % availability over seven months**,[^170] and the build-out from 4 to 24 data-center modules is under way.[^171] In Abu Dhabi, **Masdar** is building the world's first gigawatt 24/7 solar plant, because **5.2 GW of photovoltaics plus 19 GWh of battery**[^172] (storage from BYD and Sungrow of China)[^173] will deliver a constant 1 GW from 2027, a 6.1 bn $ investment (financial close July 2026)[^174] on an area roughly the size of Manhattan. The buffering itself is conceptually simple and is called **overbuild plus storage**. During the day the roughly fivefold-oversized field covers the load *and* charges the battery, at night storage takes over; 19 GWh corresponds to \~19 full-load hours at 1 GW and buffers the \~12-hour night with a clear reserve for overcast days. In sunny regions this sustains round the clock, and solar plus battery is at the same time the plant option **without a queue**, because even Masdar's gigawatt scale takes only around two years from groundbreaking to operation, while a turbine has four years of ordering lead time.[^175] (Where solar delivers around the clock even *without* a battery is shown by the next section, namely in orbit.)

**The bottleneck chain also has a bypass.** Between power plant and data center sit further bottlenecks. **Large transformers** have on average well over two years' lead time, individual orders now over 5 years, and the US imports \~80 %;[^176][^177] on top of that come switchgear, **HVDC cables** for high-voltage direct-current transmission (the three dominant manufacturers NKT/Denmark with a large high-voltage plant in Cologne, Prysmian/Italy and Nexans/France have order books for over a decade)[^178] and the interconnection queues themselves. European manufacturers sit deep in this chain too. Siemens Energy builds large transformers, **SGB-SMIT** (Regensburg) is one of Europe's large independent transformer builders,[^179] and **Maschinenfabrik Reinhausen** (also Regensburg) holds around half the world market for **tap changers**, the moving heart of every large transformer (manufacturer figure).[^180] Physically the transformer cannot be conjured away, but the grid can be bypassed, and four paths became established in 2025/26:

1. **The first path generates behind the meter ("behind the meter").** Own turbines on the campus feed in at medium voltage, and this bypasses the transmission grid, large transformers and the queue all at once. xAI runs up to 35 mobile gas turbines in Memphis (15 are permitted, including lawsuits over air quality),[^181] 41 permanent ones are approved for Colossus 2,[^182] and Stargate Abilene is building over 1 GW of own generation.[^157]
2. **The second path reactivates brownfields with an existing connection.** Decommissioned power-plant and industrial sites already have a grid connection, transformers and switchgear in place. The largest example is **Homer City** (Pennsylvania), where the state's formerly largest coal plant is being turned into a 4.5 GW gas-fired data-center campus, the largest gas plant in the US;[^183] the Three Mile Island restart and the conversion of crypto-mining sites follow the same logic. **Existing grid connections have thereby become the scarcest asset.**
3. **The third path is DC architectures in the data center.** Nvidia is switching the data-center power supply to **800-volt direct current** (in series from 2027 with the Kyber racks; partners from Infineon/Germany to Eaton/US and Schneider Electric/France), with fewer conversion stages, less copper and \~5 % efficiency gain (manufacturer figure).[^184]
4. **The fourth path is pure procurement tactics.** These include slot reservations years in advance and capacity expansion by manufacturers (Hitachi Energy of Switzerland and Japan is running an expansion program now at \~6 bn $).[^185][^186]

The transformer therefore remains necessary in every case, but the public grid with its queues can be bypassed.

**China practically does not know the Western equipment bottlenecks, because it is itself the largest equipment supplier.** China manufactures more than half of the world's transformers (TBEA is the No. 1; domestic lead time \~12 months instead of 2.5–5 years),[^187] and US data centers are now buying Chinese transformers in bulk themselves (exports in 2025 +36 %).[^188] It is also the only country running an **ultra-high-voltage grid** (>40,000 km;[^189] the 1,100 kV Changji-Guquan line brings 12 GW of desert electricity 3,300 km to eastern China, and that is how desert solar reaches the load centers).[^190] The last gap is **heavy-duty gas turbines** (until now licensed manufacture from GE, Mitsubishi and Ansaldo;[^191][^192] the first indigenous 300 MW class has been in commercial operation since late 2025),[^193] but strategically almost irrelevant, because gas supplies only \~2–3 % of China's electricity.[^194] China's own bottleneck lies elsewhere, because the generation build-out is outrunning even the world's largest grid expansion,[^195] and curtailment of solar and wind is rising steeply (H1 2026, official, 8.6 % and 9.1 % respectively; unofficial estimates including unreported curtailment are around \~26 %).[^196]

All these deals and workarounds buy time, and a structural solution they are not. Gas bridges, nuclear takes over later, solar flanks, and the next section shows where solar loses its only weakness, the night.

```mermaid
flowchart LR
    G["Short term · natural gas<br/><small>Crusoe 29 GE Vernova (~1 GW)<br/>Stargate Abilene (~1 GW)<br/>Turbine backlog 3–5 yrs</small>"] --> N["2027–2028 · nuclear<br/><small>AWS-Talen 1.92 GW<br/>Microsoft TMI ~835 MW</small>"]
    N --> S["Later · SMR<br/><small>Google-Kairos · small market<br/>(~6.9 bn $), too late</small>"]
    B["In parallel · solar + battery<br/><small>no queue · Nevada 99.2 %<br/>Masdar 1 GW 24/7 (2027)</small>"] --> O["End point · orbit<br/><small>Solar without night,<br/>AI satellites (2030s)</small>"]
    S --> O
    E["Physical bottleneck<br/><small>Turbines · transformers ·<br/>queues 5+ yrs</small>"] -.slows the build-out.-> G
```

## AI satellites: the solution beyond the power wall

If electricity is the binding constraint, the structural answer lies where electricity is not scarce, namely **in orbit**. Two physical arguments make AI satellites a serious solution rather than science fiction:

- **The energy in orbit is effectively unlimited.** In a **sun-synchronous dawn-dusk orbit**, which always runs along the Earth's day-night boundary, a satellite is almost continuously in the sun, with no night, no weather and no atmosphere in between. A solar panel there delivers up to **\~8× as much energy** as on the ground (Google research figure).[^197] And everything that makes up the terrestrial power wall drops away, meaning grid-connection queues, transformer lead times, permits and competition for land.
- **Cooling is simpler and needs not a liter of water.** The waste heat is radiated directly into the \~3 kelvin cold of outer space via **infrared radiators**, passively, with low maintenance, without cooling towers and without water consumption, while on Earth water is a bottleneck resource in its own right (see raw materials section). The physical price is that radiation is the *only* cooling path and the radiator surfaces have to be dimensioned accordingly.[^197]

The entry is already concrete, though still tiny. **Starcloud-1** put the first Nvidia H100 into orbit in November 2025[^198] and ran and trained an LLM there;[^199] in August 2026 another 250 mn $ of funding followed at a 2.3 bn $ valuation, with Nvidia participating,[^200] including the plan for a space-grade Rubin GPU from late 2028, which directly addresses the radiation objection.[^201] Google is researching solar-powered TPU satellites with **Project Suncatcher** (prototype mission with Planet Labs announced for \~early 2027, nothing in orbit),[^197] Nvidia showed a **Space-1 Vera Rubin module** (2027),[^202] SpaceX in June 2026 unveiled its orbital data-center satellite **"AI1"**,[^203] whose constellation is now called **"Starmind"**, with an FCC filing for up to 1 mn satellites[^204] and, according to Musk, first launches at the end of 2027 with Nvidia Rubin hardware (own figure),[^205] and Blue Origin filed with the FCC for **"Project Sunrise"** (>50,000 satellites).[^206] Europe's contribution so far is an EU-funded feasibility study (**ASCEND**, led by the Franco-Italian satellite builder Thales Alenia Space), meaning paper and no hardware.[^207]

China is moving in parallel, with hardware already in orbit. Zhejiang Lab launched the first twelve computing satellites of its **"Three-Body Computing Constellation"** on May 14, 2025 (target 100 satellites by 2027),[^208] ADAspace is planning a constellation of **2,800 computing satellites** (2,400 for inference, 400 for training, in dawn-dusk and sun-synchronous orbits)[^209][^210] and in November 2025 was the first to run a large language model (Qwen3) directly in orbit;[^211] in July 2026 a further constellation targeting 1,000 satellites launched its first batch,[^212] and Beijing has bundled more than a hundred space-computing organizations under a new committee, with a 1,000-POPS orbital supercomputer written into the five-year plan.[^213] The orbit question is thus no US solo effort but a race between both blocs.

**Between idea and solution stands a reality check.** Economically, computing in orbit today still costs \~4× as much as on Earth (falling with launch costs, tied above all to Starship),[^197] high-performance chips are not radiation-hardened, and everything at gigawatt scale is a 5–20-year vision; moreover, **launch capacity** itself is emerging as the next bottleneck (the Falcon 9 program ends in 2028, and demand for launch slots exceeds supply).[^200] The timing is honestly staged, with **demonstration today** (individual GPUs in orbit), **prototype constellations from \~2027** (Suncatcher, Space-1)[^197][^202] and **commercial scale in the 2030s**, with gigawatt-scale relief of the power wall coming more toward the end of that range. **The power wall stands on Earth, and in orbit it does not exist.**

---

## Conclusion: the physical diagnosis

**The bottleneck of the AI economy remains physical for the foreseeable future.** Anyone wanting to scale in 2027/28 is fighting for gigawatts, transformers and copper, not GPUs alone. Energy is becoming a strategic AI variable, because demand is unbounded, supply hangs on the electricity grid, and the only structural answer beyond the grid, the AI satellites, will bear fruit in the 2030s at the earliest. For anyone using AI, this means that the scarcity is real, that prices and availability remain volatile, and that the sovereignty questions (where does my model run, on whose grid, with whose chips?) are not theoretical.

The three parts of the series build on one another. Part 2 (The models of AI) deals with the software side, meaning how models come into being, who offers them and how you work with them, and Part 3 (The application of AI) fills out the map for operations, from desktop applications through robotics to your own company memory. The AI economy is serious about this, all the way into orbit, and its limit is electricity. What comes into being with this compute is the models, and they are the subject of Part 2.

[^1]: Tom's Hardware: "Google, Microsoft, Meta, and Amazon capex spending to hit $725 billion in 2026, up 77% from last year - analyst says bear thesis is 'garbage'". <https://www.tomshardware.com/tech-industry/big-tech/big-techs-ai-spending-plans-reach-725-billion> (accessed August 28, 2026). Supports: Big Tech capex for AI infrastructure \~725 bn $ in 2026 (+77 %).
[^2]: Pichai, Sundar: "I/O 2026: Welcome to the agentic Gemini era". Google, May 19, 2026. <https://blog.google/innovation-and-ai/sundar-pichai-io-2026/> (accessed August 28, 2026). Supports: token time series 9.7 tn (05/2024) → \~480 tn (05/2025) → >3,200 tn tokens/month (05/2026), self-reported.
[^3]: Das, Deedy and Murphy, Matt: "OpenRouter Now Processes More Than a Quadrillion Tokens a Year". Menlo Ventures, May 26, 2026. <https://menlovc.com/perspective/openrouter-now-processes-more-than-a-quadrillion-tokens-a-year/> (accessed August 28, 2026). Supports: OpenRouter volume \~15× in one year (investor source, based on billing data).
[^4]: Pichai, Sundar: "Q2 2026 earnings call: Remarks from our CEO". Alphabet, July 22, 2026. <https://blog.google/company-news/inside-google/message-ceo/alphabet-earnings-q2-2026/> (accessed August 28, 2026). Supports: Google's model APIs process \~22 bn tokens per minute, +38 % quarter-on-quarter (own figure).
[^5]: Yahoo Tech: "ChatGPT Tops 1B Weekly Users As OpenAI Rolls Out GPT-5.6 With Smarter Reasoning". August 6, 2026. <https://tech.yahoo.com/ai/chatgpt/articles/chatgpt-tops-1b-weekly-users-191537749.html> (not retrievable on August 28, 2026). Supports: ChatGPT tops 1 bn weekly active users.
[^6]: CNBC: "Anthropic CEO Dario Amodei says company crew 80-fold in first quarter". May 6, 2026. <https://www.cnbc.com/2026/05/06/anthropic-ceo-dario-amodei-says-company-crew-80-fold-in-first-quarter.html> (not retrievable on August 28, 2026). Supports: Anthropic grew 80× annualized in Q1 2026, "too hard to handle" (Amodei, self-reported).
[^7]: Nvidia: "NVIDIA Announces Financial Results for First Quarter Fiscal 2027". May 20, 2026. <https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-first-quarter-fiscal-2027> (accessed August 28, 2026). Supports: Q1 FY2027, data-center revenue 75.2 bn $, +92 % year-on-year.
[^8]: OpenAI: "OpenAI raises $122 billion to accelerate the next phase of AI". March 31, 2026. <https://openai.com/index/accelerating-the-next-phase-ai/> (not retrievable on August 28, 2026). Supports: funding round of 122 bn $ at an 852 bn $ post-money valuation (self-reported).
[^9]: Anthropic: "Anthropic raises $65B in Series H funding at $965B post-money valuation". May 28, 2026. <https://www.anthropic.com/news/series-h> (accessed August 28, 2026). Supports: Series H of 65 bn $ at a 965 bn $ valuation (self-reported).
[^10]: CNBC: "Anthropic confidentially files IPO prospectus with SEC, prepping Wall Street for landmark AI deal". June 1, 2026. <https://www.cnbc.com/2026/06/01/anthropic-ipo-s1-prospectus.html> (not retrievable on August 28, 2026). Supports: Anthropic confidentially files S-1 for its IPO.
[^11]: CNBC: "OpenAI confidentially files for IPO, prepping Wall Street for mega AI debut". June 8, 2026. <https://www.cnbc.com/2026/06/08/openai-confidentially-files-for-ipo-prepping-wall-street-for-ai-debut.html> (not retrievable on August 28, 2026). Supports: OpenAI confidentially files for an IPO.
[^12]: StartupHub.ai: "OpenAI Completes $7 Billion Employee Tender Offer at $852 Billion Valuation". August 11, 2026. <https://www.startuphub.ai/ai-news/ipo-watch/2026/openai-7b-tender-offer-2026-08-11> (accessed August 28, 2026). Supports: OpenAI tender offer of \~7 bn $ at 852 bn $, leaning toward an IPO only in 2027.
[^13]: Bort, Julie: "Cerebras raises $5.5B, then stock pops 108%, in the first huge tech IPO of 2026". TechCrunch, May 14, 2026. <https://techcrunch.com/2026/05/14/cerebras-raises-5-5b-kicking-off-2026s-ipo-season-with-a-bang/> (accessed August 28, 2026). Supports: Cerebras IPO, proceeds 5.55 bn $, +68 % on day one.
[^14]: CNBC: "Nvidia-backed CoreWeave closes flat at $40 after biggest U.S. tech IPO since 2021". March 28, 2025. <https://www.cnbc.com/2025/03/28/coreweave-starts-trading-on-nasdaq-at-per-share.html> (not retrievable on August 28, 2026). Supports: CoreWeave Nasdaq debut (1.5 bn $ proceeds at \~23 bn $ valuation) as the first major AI-infrastructure IPO.
[^15]: CNBC: "MiniMax doubles in Hong Kong debut, marking yet another Chinese AI listing". January 9, 2026. <https://www.cnbc.com/2026/01/09/minimax-hong-kong-ipo-ai-tigers-zhipu.html> (not retrievable on August 28, 2026). Supports: MiniMax doubles in Hong Kong debut (proceeds \~620 mn $), one day after Zhipu's listing.
[^16]: Bloomberg: "Zhipu Considers Multibillion-Dollar Share Sale in Hong Kong After 2,000% Gain". June 24, 2026. <https://www.bloomberg.com/news/articles/2026-06-24/zhipu-considers-multibillion-dollar-share-sale-in-hong-kong-after-2-000-gain> (not retrievable on August 28, 2026). Supports: Zhipu considers multibillion-dollar share sale after \~2,000 % gain since its January debut.
[^17]: Han, Yuhang: "Chinese Memory-Chip Maker CXMT Sets Subscription Date for Shanghai IPO". Caixin, July 9, 2026. <https://www.caixinglobal.com/2026-07-09/chinese-memory-chip-maker-cxmt-sets-subscription-date-for-shanghai-ipo-102462287.html> (accessed August 28, 2026). Supports: CXMT sets subscription date for its Shanghai IPO, proceeds target.
[^18]: CNBC: "Chipmaker CXMT’s 466% market debut surge makes it the most valuable China-listed company". July 27, 2026. <https://www.cnbc.com/2026/07/27/cxmt-china-market-debut-chipmaker-ipo.html> (not retrievable on August 28, 2026). Supports: CXMT debut +466 %, proceeds \~8.6 bn $, largest China chip IPO since SMIC.
[^19]: Du, Zhihang: "Unitree Robotics Wins Approval for $618 Million STAR Market IPO". Caixin, July 3, 2026. <https://www.caixinglobal.com/2026-07-03/unitree-robotics-wins-approval-for-618-million-star-market-ipo-102460136.html> (accessed August 28, 2026). Supports: Unitree wins approval for STAR Market IPO.
[^20]: Bloomberg: "Unitree Robotics Surges 460% After $904 Million Shanghai IPO". August 18, 2026. <https://www.bloomberg.com/news/articles/2026-08-18/unitree-robotics-set-to-debut-after-904-million-shanghai-ipo> (not retrievable on August 28, 2026). Supports: Unitree debut after 904 mn $ Shanghai IPO (+460 % on day one).
[^21]: Estrada, Sheryl: "Why the 2026 IPO boom is about to broaden beyond AI mega-deals". Fortune, July 11, 2026. <https://fortune.com/2026/07/11/why-2026-ipo-boom-broaden-beyond-ai-mega-deals/> (accessed August 28, 2026). Supports: 2026 on track to become the largest IPO year of all time, driven by AI.
[^22]: Anani, Karim: "EY Global IPO Trends Q2 2026". EY, July 7, 2026. <https://www.ey.com/en_us/insights/ipo/trends> (accessed August 28, 2026). Supports: US IPOs H1 2026, twelve deals above 1 bn $, the first half alone would rank as the second most active full year by proceeds; AI as the main driver.
[^23]: Dufresne, Alexis: "US Venture Hits $412.7B in H1 2026, AI Takes 86%". AI Weekly, July 9, 2026. <https://aiweekly.co/alerts/us-venture-hits-4127b-in-h1-2026-ai-takes-86> (accessed August 28, 2026). Supports: \~86–87 % of US venture dollars go to AI.
[^24]: Rogelberg, Sasha: "Satya Nadella believes Chinese AI startup DeepSeek could be a win for tech, even as Microsoft's shares tumble". Fortune, January 27, 2025. <https://fortune.com/2025/01/27/microsoft-ceo-satya-nadella-deepseek-optimism-jevons-paradox/> (accessed August 28, 2026). Supports: Nadella quote "Jevons paradox strikes again!" after the DeepSeek shock.
[^25]: Cottier, Ben and Snodin, Ben et al.: "LLM inference prices have fallen rapidly but unequally across tasks". Epoch AI, March 12, 2025. <https://epoch.ai/data-insights/llm-inference-price-trends> (accessed August 28, 2026). Supports: price per unit of intelligence falls 9–900× per year, GPT-4 level >40× cheaper in two years.
[^26]: Jiang, Ben: "After triggering price war, DeepSeek reverses course with surcharge on peak-hour API use". SCMP, June 30, 2026. <https://www.scmp.com/tech/big-tech/article/3358868/after-triggering-price-war-deepseek-reverses-course-surcharge-peak-hour-api-use> (accessed August 28, 2026). Supports: DeepSeek price cut of up to −90 % → +297 % token volume on day one → peak-hour surcharges.
[^27]: Moreau, Clément: "DeepSeek Price Shake-Up: How Pricing Volatility Changes AI API Economics". Eden AI, August 13, 2026. <https://www.edenai.co/post/deepseek-price-shake-up-how-pricing-volatility-changes-ai-api-economics> (accessed August 28, 2026). Supports: DeepSeek announces price increases for the V4 generation.
[^28]: Emberson, Luke and Cottier, Ben et al.: "LLM responses to benchmark questions are getting longer over time". Epoch AI, April 17, 2025. <https://epoch.ai/data-insights/output-length> (accessed August 28, 2026). Supports: reasoning models produce on average \~8× more tokens per response than non-reasoning models, response length growing \~5×/year.
[^29]: Bai, Longju and Huang, Zhemin et al.: "How Do AI Agents Spend Your Money? Analyzing and Predicting Token Consumption in Agentic Coding Tasks". arXiv, April 24, 2026. <https://arxiv.org/abs/2604.22750> (accessed August 28, 2026). Supports: agentic tasks consume up to \~1,000× more tokens than chat (measurement; 10–100× in practice).
[^30]: Liu, Phoebe: "OpenAI Could Be Blowing As Much As $15 Million Per Day On Silly Sora Videos". Forbes, November 10, 2025. <https://www.forbes.com/sites/phoebeliu/2025/11/10/openai-spending-ai-generated-sora-videos/> (not retrievable on August 28, 2026). Supports: Sora burns \~15 mn $ per day (estimate).
[^31]: Chui, Michael and Hazan, Eric et al.: "The economic potential of generative AI: The next productivity frontier". McKinsey, June 14, 2023. <https://www.mckinsey.com/capabilities/tech-and-ai/our-insights/the-economic-potential-of-generative-ai-the-next-productivity-frontier> (not retrievable on August 28, 2026). Supports: value potential of generative AI 2.6–4.4 tn $ per year (model estimate).
[^32]: Tom's Hardware: "OpenAI has run out of GPUs, says Sam Altman - GPT-4.5 rollout delayed due to lack of processing power". February 2025. <https://www.tomshardware.com/tech-industry/artificial-intelligence/openai-has-run-out-of-gpus-says-sam-altman-gpt-4-5-rollout-delayed-due-to-lack-of-processing-power> (accessed August 28, 2026). Supports: Altman "out of GPUs", GPT-4.5 rollout delayed for lack of compute.
[^33]: Gerut, Amanda: "Microsoft, Meta, and Google just announced billions more in AI spending. Only Google convinced investors it's paying off". Fortune, April 29, 2026. <https://fortune.com/2026/04/29/microsoft-meta-google-ai-capex-spending-billions/> (accessed August 28, 2026). Supports: Microsoft CFO Hood capacity-constrained "at least through 2026" despite \~190 bn $ capex; Pichai "compute constrained".
[^34]: Sparks, Daniel: "Jensen Huang Used 1 Word to Describe AI Demand. It Could Be the Most Important of 2026.". The Motley Fool, May 30, 2026. <https://www.fool.com/investing/2026/05/30/jensen-huang-used-1-word-to-describe-ai-demand-it/> (accessed August 28, 2026). Supports: Huang "demand has gone parabolic" (beneficiary's PR).
[^35]: Nvidia: "H200 GPU". <https://www.nvidia.com/en-us/data-center/h200/> (accessed August 28, 2026). Supports: memory/bandwidth series A100 through B300 (manufacturer figures).
[^36]: JEDEC: "JEDEC® and Industry Leaders Collaborate to Release JESD270-4 HBM4 Standard: Advancing Bandwidth, Efficiency, and Capacity for AI and HPC". April 16, 2025. <https://www.jedec.org/news/pressreleases/jedec-publishes-hbm4-standard> (not retrievable on August 28, 2026). Supports: HBM4 standard JESD270-4, 2,048-bit interface, up to 36 GB per stack.
[^37]: SK Hynix: "SK hynix Completes World's First HBM4 Development and Readies Mass Production". September 12, 2025. <https://news.skhynix.com/sk-hynix-completes-worlds-first-hbm4-development-and-readies-mass-production/> (accessed August 28, 2026). Supports: HBM4 development completed, mass production readied (self-reported).
[^38]: Counterpoint Research: "Global DRAM and HBM Market Share: Quarterly". June 8, 2026. <https://counterpointresearch.com/en/insights/global-dram-and-hbm-market-share> (accessed August 28, 2026). Supports: HBM market shares SK Hynix \~58 %, Samsung \~21 %, Micron \~21 % (analyst estimate).
[^39]: TrendForce: "[News] Samsung's HBM4 Yield Reportedly Hits 80% as HBM Race Heats Up, SK hynix Labor Talks Add a Twist". August 10, 2026. <https://www.trendforce.com/news/2026/08/10/news-samsungs-hbm4-yield-reportedly-hits-80-as-race-to-supply-vera-rubin-heats-up-sk-hynix-labor-talks-add-a-twist/> (accessed August 28, 2026). Supports: Samsung as first HBM4 supplier for Nvidia's Vera Rubin, HBM4 yield reportedly \~80 %.
[^40]: Blau, John: "UPDATE: Germany's DRAM Bailout Hits a Snag as Qimonda Goes Bankrupt". IEEE Spectrum, January 28, 2009. <https://spectrum.ieee.org/update-germanys-dram-bailout-hits-a-snag-as-qimonda-goes-bankrupt> (accessed August 28, 2026). Supports: Qimonda bankruptcy, the end of the last European volume DRAM manufacturer.
[^41]: Robinson, Cliff: "This is the Google TPU v7 Ironwood Chip". ServeTheHome, November 20, 2025. <https://www.servethehome.com/this-is-the-google-tpu-v7-ironwood-chip/> (accessed August 28, 2026). Supports: Google TPU v7 "Ironwood", chip details.
[^42]: Patel, Dylan and Xie, Myron et al.: "TPUv7: Google Takes a Swing at the King". SemiAnalysis, November 28, 2025. <https://newsletter.semianalysis.com/p/tpuv7-google-takes-a-swing-at-the> (accessed August 28, 2026). Supports: TPUv7 Ironwood, first inference-optimized TPU (analysis).
[^43]: AMD: "AMD and OpenAI Announce Strategic Partnership to Deploy 6 Gigawatts of AMD GPUs". October 6, 2025. <https://www.amd.com/en/newsroom/press-releases/2025-10-6-amd-and-openai-announce-strategic-partnership-to-d.html> (not retrievable on August 28, 2026). Supports: partnership with OpenAI for 6 GW of AMD GPUs, first 1 GW of MI450 from H2 2026.
[^44]: Nvidia: "Rack-Scale Agentic AI Supercomputer | NVIDIA Vera Rubin NVL72". <https://www.nvidia.com/en-us/data-center/vera-rubin-nvl72/> (accessed August 28, 2026). Supports: Rubin GPU \~1,800–2,300 W, 336 bn transistors; NVL72 with 72 GPUs + 36 CPUs, 20.7 TB HBM4, 3.6 EFLOPS (manufacturer figure).
[^45]: Eliahou Ontiveros, Jeremie and Patel, Dylan et al.: "xAI's Colossus 2 - First Gigawatt Datacenter In The World, Unique RL Methodology, Capital Raise". SemiAnalysis, September 16, 2025. <https://newsletter.semianalysis.com/p/xais-colossus-2-first-gigawatt-datacenter> (accessed August 28, 2026). Supports: xAI Colossus 2 as the first gigawatt data center (analysis).
[^46]: Crosley, Blake: "xAI Colossus Hits 2 GW: 555,000 GPUs, $18B, Largest AI Site". Introl, January 3, 2026. <https://introl.com/blog/xai-colossus-2-gigawatt-expansion-555k-gpus-january-2026> (accessed August 28, 2026). Supports: Colossus 2, \~555,000 GPUs under construction, on track for \~1 mn GPUs / 2 GW (xAI self-reported).
[^47]: DCD: "OpenAI and Oracle to deploy 450,000 GB200 GPUs at Stargate data center in Abilene, Texas". <https://www.datacenterdynamics.com/en/news/openai-and-oracle-to-deploy-450000-gb200-gpus-at-stargate-abilene-data-center/> (not retrievable on August 28, 2026). Supports: OpenAI/Oracle, >450,000 GB200 GPUs at Stargate Abilene (1.2 GW).
[^48]: OpenAI: "OpenAI, Oracle, and SoftBank expand Stargate with five new AI data center sites". September 23, 2025. <https://openai.com/index/five-new-stargate-sites/> (not retrievable on August 28, 2026). Supports: five new Stargate sites, program \~10 GW by 2029 (self-reported).
[^49]: EuroHPC JU: "JUPITER: Launching Europe’s Exascale Era". September 5, 2025. <https://www.eurohpc-ju.europa.eu/jupiter-launching-europes-exascale-era-2025-09-05_en> (accessed August 28, 2026). Supports: JUPITER (Jülich), Europe's first exascale computer, \~24,000 Nvidia GH200 superchips.
[^50]: European Commission: "AI Gigafactories". <https://commission.europa.eu/topics/competitiveness/competitiveness-coordination-tool-projects/ai-gigafactories_en> (accessed August 28, 2026). Supports: AI Gigafactories (InvestAI, \~100,000 GPUs per site), formal call summer 2026, no construction start.
[^51]: OpenEuroLLM: "Open LLMs for Transparent AI in Europe". February 3, 2025. <https://openeurollm.eu/blog/launch-press-release> (accessed August 28, 2026). Supports: EU model initiative, so far without a competitive model.
[^52]: Steinschaden, Jakob: "Cohere aus Kanada übernimmt Aleph Alpha, gemeinsame Bewertung bei 20 Milliarden Dollar". Trending Topics, April 24, 2026. <https://www.trendingtopics.eu/aleph-alpha-und-cohere-aus-kanada-vor-fusion-bewertung-bei-etwa-20-milliarden-dollar/> (accessed August 28, 2026). Supports: Aleph Alpha ahead of takeover by Cohere (valuation \~20 bn $).
[^53]: LaPedus, Mark: "TSMC Gains Foundry Share in Q1 ’26". Substack, June 12, 2026. <https://marklapedus.substack.com/p/tsmc-gains-foundry-share-in-q1-26> (accessed August 28, 2026). Supports: TSMC foundry share \~72 % overall, >90 % leading-edge (analyst estimate).
[^54]: Zeiss SMT: "EUV lithography and technology". <https://www.zeiss.com/semiconductor-manufacturing-technology/inspiring-technology/euv-lithography.html> (accessed August 28, 2026). Supports: complete EUV mirror optics; precision comparison "Germany-sized mirror, \~0.1 mm unevenness" (manufacturer figure).
[^55]: Trumpf: "Generation of EUV radiation". <https://www.trumpf.com/en_US/solutions/applications/euv-lithography/> (accessed August 28, 2026). Supports: EUV light source, high-power CO2 laser, 50,000 tin droplets per second (manufacturer figure).
[^56]: ASML: "Berlin, Germany - ASML Berlin". <https://www.asml.com/en/company/about-asml/locations/berlin> (accessed August 28, 2026). Supports: former Berliner Glas, acquired 2020, >1,700 employees, wafer stages, reticle chucks, mirror blocks (manufacturer figure).
[^57]: NineScrolls: "Imec Installs $400M ASML EXE:5200 High-NA EUV - One of Fewer Than 12 Worldwide - Targeting Sub-2nm by Q4 2026". April 1, 2026. <https://ninescrolls.com/news/imec-installs-400m-asml-exe-5200-high-na-euv-one-of-fewer-than-12-worldwide-targ> (accessed August 28, 2026). Supports: imec installs ASML EXE:5200 High-NA EUV (\~400 mn $), fewer than twelve worldwide.
[^58]: TSMC: "TSMC Intends to Expand Its Investment in the United States to US$165 Billion to Power the Future of AI". March 4, 2025. <https://pr.tsmc.com/english/news/3210> (accessed August 28, 2026). Supports: Arizona investment expanded to 165 bn $, Gigafab cluster with six fabs and two packaging plants.
[^59]: TSMC: "TSMC, Bosch, Infineon, and NXP Establish Joint Venture to Bring Advanced Semiconductor Manufacturing to Europe". August 8, 2023. <https://pr.tsmc.com/english/news/3049> (accessed August 28, 2026). Supports: ESMC Dresden, TSMC 70 % + Bosch/Infineon/NXP 10 % each, >10 bn €, 28/22 and 16/12 nm processes, production start end-2027 (manufacturer figure).
[^60]: Mantel, Mark: "Intel gives up Magdeburg fab and raises the prospect of the end of the foundry". heise, July 25, 2025. <https://www.heise.de/en/news/Intel-gives-up-Magdeburg-fab-and-announces-end-of-foundry-10499170.html> (accessed August 28, 2026). Supports: Intel definitively gives up the Magdeburg fab.
[^61]: van Gerven, Paul: "EU Chips Act “needs a reality check,” say auditors". Bits&Chips, April 29, 2025. <https://bits-chips.com/article/eu-chips-act-needs-a-reality-check-say-auditors/> (accessed August 28, 2026). Supports: European Court of Auditors, Chips Act target of 20 % world market share by 2030 "very unlikely" (Commission projection 11.7 %), >43 bn € mobilized.
[^62]: Markman, Jon: "Intel Joins Terafab To Build Elon Musk’s $25 Billion AI Chip Project". Forbes, April 10, 2026. <https://www.forbes.com/sites/jonmarkman/2026/04/10/intel-joins-terafab-to-build-elon-musks-25b-ai-chip-project/> (accessed August 28, 2026). Supports: Intel joins Musk's Terafab project (\~25 bn $) as foundry partner.
[^63]: Tech-Insider: "Intel Joins Musk’s $25B Terafab Chip Factory [2026]". <https://tech-insider.org/intel-terafab-foundry-deal-musk-chip-manufacturing-2026/> (not retrievable on August 28, 2026). Supports: Intel-Terafab foundry deal, chip manufacturing project of Tesla/SpaceX/xAI.
[^64]: Picchi, Aimee: "What is Elon Musk’s Terafab chip project? Here are his “most epic” goals for the factory.". CBS News, March 23, 2026. <https://www.cbsnews.com/news/terafab-elon-musk-chips-semiconductors-what-to-know/> (accessed August 28, 2026). Supports: Terafab, Musk's chip project targeting "1 terawatt of AI compute per year"; phase 1 Grimes County, Texas, 16.8 bn $ (announcement).
[^65]: Micron: "Micron in High-Volume Production of HBM4 Designed for NVIDIA Vera Rubin, PCIe Gen6 SSD and SOCAMM2". March 16, 2026. <https://investors.micron.com/news-releases/news-release-details/micron-high-volume-production-hbm4-designed-nvidia-vera-rubin> (accessed August 28, 2026). Supports: HBM4 high-volume production for Nvidia Vera Rubin; HBM sold out for the calendar year (self-reported).
[^66]: Mellor, Chris: "High-bandwidth memory v4 supply takes shape". Blocks & Files, January 28, 2026. <https://www.blocksandfiles.com/hci/2026/01/28/high-bandwidth-memory-v4-supply-takes-shape/4090314> (accessed August 28, 2026). Supports: HBM4 supply, Rubin output in 2026 possibly capped at \~200–300 k GPUs (supply-chain estimate).
[^67]: Crosley, Blake: "NVIDIA Rubin Enters Full Production: The 336 Billion Transistor GPU Reshaping AI Infrastructure". Introl, January 8, 2026. <https://introl.com/blog/nvidia-rubin-full-production-ces-2026-ai-infrastructure> (accessed August 28, 2026). Supports: Nvidia Vera Rubin in full production (CES 2026, manufacturer figures).
[^68]: TrendForce: "[News] TSMC CoWoS Supply-Demand Gap Reportedly Seen Narrowing from 20% to 10% by End-2026". June 15, 2026. <https://www.trendforce.com/news/2026/06/15/news-tsmc-cowos-supply-demand-gap-reportedly-seen-narrowing-from-20-to-10-by-end-2026-as-capacity-expands/> (accessed August 28, 2026). Supports: CoWoS supply-demand gap reportedly narrowing from \~20 % to \~10 % by end-2026; TSMC capacity 120–140 k wafers/month.
[^69]: TSMC: "Q1 2026 Taiwan Semiconductor Manufacturing Co Ltd Earnings Call". April 16, 2026. <https://investor.tsmc.com/english/encrypt/files/encrypt_file/reports/2026-04/3cef85204275f94fd111485cfdf4adb3c0263c45/TSMC%201Q26%20Transcript.pdf> (accessed August 28, 2026). Supports: C.C. Wei, supply remains "very tight", including 2027; three new N3 fabs announced.
[^70]: CSIS: [Title not determinable]. <https://www.csis.org/analysis/russia-ukraine-war-semiconductor-supply-chains> (not retrievable on August 28, 2026). Supports: neon, Ukraine supplied \~90 % of US semiconductor demand, diversified since 2022.
[^71]: Reuters: [Title not determinable]. October 1, 2024. <https://www.reuters.com/markets/commodities/quartz-mines-hurricane-helene-spruce-pine-2024-10-01/> (not retrievable on August 28, 2026). Supports: Hurricane Helene shuts down the Spruce Pine quartz mines (\~70–90 % of world supply).
[^72]: Qin, Min and Han, Wei: "China Piles $47.5 Billion Into ‘Big Fund III’ to Boost Chip Development". Caixin, May 28, 2024. <https://www.caixinglobal.com/2024-05-28/china-piles-475-billion-into-big-fund-iii-to-boost-chip-development-102200633.html> (accessed August 28, 2026). Supports: Big Fund III, 344 bn yuan ≈ 47.5 bn $, larger than both predecessors combined.
[^73]: European Commission: "Proposal for the Chips Act 2.0". June 3, 2026. <https://digital-strategy.ec.europa.eu/en/library/proposal-chips-act-20> (accessed August 28, 2026). Supports: proposal "Chips Act 2.0", for the first time with a leading-edge fab target.
[^74]: Ahmad, Afzal and Wagner, Andrew: "Is SMIC N+3’s Metal Pitch Smaller than Intel 18A’s?". SemiAnalysis, June 14, 2026. <https://newsletter.semianalysis.com/p/steel-smic-n3-teardown> (accessed August 28, 2026). Supports: "5 nm" reports really \~TSMC N6 class, yields estimated 20–60 %, wafer costs +40–50 % (analyst estimate).
[^75]: Washenko, Anna: "China reportedly has a prototype EUV machine built by ex-ASML employees". Engadget, December 17, 2025. <https://www.engadget.com/big-tech/china-reportedly-has-a-prototype-euv-machine-built-by-ex-asml-employees-235833756.html> (accessed August 28, 2026). Supports: China's EUV prototype in Shenzhen, built by ex-ASML employees, no chip exposed yet.
[^76]: The Diplomat: "China’s EUV Lithography Progress: Parsing Signal From Noise". July 8, 2026. <https://thediplomat.com/2026/07/chinas-euv-lithography-progress-parsing-signal-from-noise/> (not retrievable on August 28, 2026). Supports: China's EUV progress, serial readiness in \~2030 at the earliest.
[^77]: CNBC: "China’s reported chip breakthrough comes with some big caveats". July 28, 2026. <https://www.cnbc.com/2026/07/28/china-chipmaking-duv-tool-asml-explained.html> (not retrievable on August 28, 2026). Supports: China's own immersion DUV machines in small series (\~5 units 2026, \~20 planned 2027), \~four generations behind ASML; 95 DUV imports in 2025.
[^78]: Tom's Hardware: "ASML CEO says China is 10 to 15 years behind in chipmaking capabilities". December 25, 2024. <https://www.tomshardware.com/tech-industry/asml-ceo-says-china-is-10-to-15-years-behind-in-chipmaking-capabilities> (accessed August 28, 2026). Supports: ASML CEO Fouquet, without EUV, China lags 10–15 years behind (NRC interview).
[^79]: Tsai, Jessica: "CXMT HBM3 timeline slips, mass production unlikely in 2026". DigiTimes, April 22, 2026. <https://www.digitimes.com/news/a20260421PD230/cxmt-hbm3-dram-production-2026.html> (accessed August 28, 2026). Supports: CXMT masters HBM2, misses its 2026 HBM3 target, \~3–4 years behind SK Hynix (analyst estimate).
[^80]: SemiAnalysis: [Title not determinable]. September 25, 2025. <https://semianalysis.com/2025/09/25/huawei-ascend-hbm-stockpile/> (not retrievable on August 28, 2026). Supports: Huawei hoards \~13 mn HBM stacks (material for \~1.6 mn Ascend), domestic production 2026 only \~250–300 k (estimate).
[^81]: McGuire, Chris: "China’s AI Chip Deficit: Why Huawei Can’t Catch Nvidia and U.S. Export Controls Should Remain". CFR. <https://www.cfr.org/articles/chinas-ai-chip-deficit-why-huawei-cant-catch-nvidia-and-u-s-export-controls-should-remain> (not retrievable on August 28, 2026). Supports: China's AI chip deficit, HBM and packaging as the bottleneck, why Huawei can't catch Nvidia.
[^82]: Patel, Dylan and Xie, Myron et al.: "Huawei Ascend Production Ramp: Die Banks, TSMC Continued Production, HBM is The Bottleneck". SemiAnalysis, September 8, 2025. <https://newsletter.semianalysis.com/p/huawei-ascend-production-ramp> (accessed August 28, 2026). Supports: Huawei Ascend production ramp and cluster scaling (Atlas 950) (analyst estimate).
[^83]: DigiTimes: "Naura rises to No. 5 in global semiconductor equipment rankings". February 13, 2026. <https://www.digitimes.com/news/a20260213PD212/naura-technology-ic-manufacturing-equipment-localization-2025.html> (accessed August 28, 2026). Supports: Naura world No. 5 in chip equipment, localization \~35 %, target 70 % by 2027 (analyst estimate).
[^84]: Kaur, Dashveenjit: "China’s chip self-sufficiency push is real this time - but the target has been here before". TechWire Asia, May 6, 2026. <https://techwireasia.com/2026/05/china-semiconductor-self-sufficiency-wafer-target-2026/> (accessed August 28, 2026). Supports: 70 % self-sufficiency target for 2025 missed by a significant margin, broad self-sufficiency \~50 % per TrendForce.
[^85]: Shen, Jessie: "China to fall far short of IC self-sufficiency goal by 2025, says IC Insights". DigiTimes, January 7, 2021. <https://www.digitimes.com/news/a20210107PR200.html> (accessed August 28, 2026). Supports: IC Insights forecasts China's IC self-sufficiency at only 19.4 % in 2025 (15.9 % in 2020).
[^86]: Trivium China: "Beijing bans foreign AI chips in state-funded data centers". November 6, 2025. <https://triviumchina.com/2025/11/06/beijing-bans-foreign-ai-chips-in-state-funded-data-centers/> (not retrievable on August 28, 2026). Supports: Beijing bans foreign AI chips in state-funded data centers.
[^87]: Tech-Insider: "Nvidia H200 China Sales: 75K Cap + 25% Tax [April 2026]". April 2026. <https://tech-insider.org/nvidia-h200-chip-sales-china-2026/> (not retrievable on August 28, 2026). Supports: H200 blocked by Chinese customs despite US release, Nvidia's China data-center revenue effectively zero.
[^88]: Pilz, Konstantin F. and Rahman, Robi et al.: "The US hosts the majority of GPU cluster performance, followed by China". Epoch AI, June 5, 2025. <https://epoch.ai/data-insights/ai-supercomputers-performance-share-by-country> (accessed August 28, 2026). Supports: AI supercomputer performance share by country, US 74.5 %, China 14.1 %, EU-27 4.8 %.
[^89]: Pilz, Konstantin F. and Sanders, James et al.: "Trends in AI Supercomputers". arXiv, April 22, 2025. <https://arxiv.org/abs/2504.16026> (accessed August 28, 2026). Supports: dataset captures only 10–20 % of global capacity.
[^90]: Omdia: "New Omdia forecast: AI data center chip market to hit $286bn, growth likely peaking as custom ASICs gain ground". August 2025. <https://omdia.tech.informa.com/pr/2025/aug/ai-data-center-chip-market-to-hit-286bn-growth-likely-peaking-as-custom-asics-gain-ground> (not retrievable on August 28, 2026). Supports: Nvidia ships \~5 mn AI accelerators in 2025 (analyst estimate).
[^91]: CNBC: "Jensen Huang says Nvidia’s AI chips are now being manufactured in Arizona". October 28, 2025. <https://www.cnbc.com/2025/10/28/nvidia-jensen-huang-gtc-washington-dc-ai.html> (not retrievable on August 28, 2026). Supports: Huang at GTC, "6 mn Blackwell GPUs in four quarters" (self-reported, counts dies).
[^92]: CFR: [Title not determinable]. <https://www.cfr.org/blog/chinas-ai-chip-gap-growing-not-shrinking> (not retrievable on August 28, 2026). Supports: China's AI chip gap is growing, Huawei's compute output \~4 % of Nvidia in 2026 (calculation).
[^93]: gov.cn: "Renewables account for over 60 pct of China’s power capacity in 2025". January 30, 2026. <https://english.www.gov.cn/archive/statistics/202601/30/content_WS697cb463c6d00ca5f9a08da7.html> (accessed August 28, 2026). Supports: China 3.89 TW of installed generation capacity at end-2025, +16 % (official).
[^94]: Ember: "Global Electricity Review 2026". April 2026. <https://ember-energy.org/latest-insights/global-electricity-review-2026/2025-in-review/> (not retrievable on August 28, 2026). Supports: China's solar additions 378 GW (DC) = 58 % of world additions in 2025.
[^95]: EIA: "U.S. electricity generation in 2025 hit a record, again". March 5, 2026. <https://www.eia.gov/todayinenergy/detail.php?id=67284> (accessed August 28, 2026). Supports: US generation capacity \~1.25–1.3 TW; 53 GW added in 2025, predominantly solar plus battery (official).
[^96]: Eurostat: "Electricity production capacities". <https://ec.europa.eu/eurostat/databrowser/view/nrg_inf_epc/default/table?lang=en> (accessed August 28, 2026). Supports: EU-27 generation capacity \~1.17 TW at end-2024 (official).
[^97]: Liu, Aixin and Feng, Bei et al.: "DeepSeek-V3 Technical Report". arXiv, December 27, 2024. <https://arxiv.org/abs/2412.19437> (accessed August 28, 2026). Supports: 2,048 H800, 2.788 mn GPU hours ≈ 5.6 mn $ final run, 671/37 bn parameters, FP8 + MLA (self-reported).
[^98]: Patel, Dylan and Knuhtsen, Reyk et al.: "DeepSeek Debates: Chinese Leadership On Cost, True Training Cost, Closed Model Margin Impacts". SemiAnalysis, January 31, 2025. <https://newsletter.semianalysis.com/p/deepseek-debates> (accessed August 28, 2026). Supports: DeepSeek/High-Flyer, access to \~50,000 Hopper GPUs (H800, H100, H20), \~1.6 bn $ server investment (analyst estimate).
[^99]: DeepSeek: "DeepSeek V4 Preview Release". April 24, 2026. <https://api-docs.deepseek.com/news/news260424/> (accessed August 28, 2026). Supports: V4 announcement, V4 Flash with 284/13 bn parameters (self-reported).
[^100]: CNBC: "Nvidia AI chips smuggled into China after Trump restrictions: Report". July 24, 2025. <https://www.cnbc.com/2025/07/24/nvidia-ai-chips-smuggling-china-trump.html> (not retrievable on August 28, 2026). Supports: FT reporting, \~1 bn $ of Nvidia chips smuggled into China in three months.
[^101]: CNBC: "Singapore, U.S. investigate Nvidia client Megaspeed". October 10, 2025. <https://www.cnbc.com/2025/10/10/singapore-us-investigate-nvidia-client-megaspeed-export-controls-violation.html> (not retrievable on August 28, 2026). Supports: Singapore and US investigate Nvidia client Megaspeed (\~2 bn $, cloud access via Malaysia/Indonesia).
[^102]: Brewster, Thomas: "AI Founder Illegally Shipped Nvidia Chips To China In $4 Million Scheme, DOJ Alleges". Forbes, November 20, 2025. <https://www.forbes.com/sites/thomasbrewster/2025/11/20/illegal-nvidia-ai-chip-sales-to-china-doj-indictment/> (accessed August 28, 2026). Supports: DOJ indictment over illegal Nvidia AI chip sales to China.
[^103]: TrendForce: "[News] DeepSeek R2 Model Launch Reportedly Delayed Amid Huawei Ascend Chip Hurdles". August 14, 2025. <https://www.trendforce.com/news/2025/08/14/news-deepseek-r2-model-launch-reportedly-delayed-amid-huawei-ascend-chip-hurdles/> (accessed August 28, 2026). Supports: DeepSeek R2 training on Huawei Ascend failed, back onto Nvidia (anonymous sources).
[^104]: Hong, Jinju: "Z.ai unveils GLM-5.2 trained with only Huawei chips, without Nvidia". Digital Today, June 19, 2026. <https://www.digitaltoday.co.kr/en/view/68873/z-ai-unveils-glm-5-2-trained-with-only-huawei-chips-without-nvidia> (accessed August 28, 2026). Supports: Z.ai/Zhipu, GLM-5.2 trained exclusively on Huawei chips (self-reported, unverified).
[^105]: Bloomberg: "Huawei Used TSMC, Samsung, SK Hynix Components in Top AI Chips". October 3, 2025. <https://www.bloomberg.com/news/articles/2025-10-03/huawei-used-tsmc-samsung-sk-hynix-components-in-top-ai-chips> (not retrievable on August 28, 2026). Supports: TechInsights teardowns find TSMC dies and Samsung/SK Hynix HBM in Ascend 910C samples (measured); \~2.9 mn dies via Sophgo (analyst estimate).
[^106]: SemiWiki: "TechInsights Teardown: Huawei Ascend 910c Still Contains CPU Dies from TSMC from 2020". October 4, 2025. <https://semiwiki.com/forum/threads/techinsights-teardown-huawei-ascend-910c-still-contains-cpu-dies-from-tsmc-from-2020.23737/> (accessed August 28, 2026). Supports: Ascend 910C still contains TSMC CPU dies from \~2020.
[^107]: US Congress: "S.4281 - Multilateral Alignment of Technology Controls on Hardware (MATCH) Act". <https://www.congress.gov/bill/119th-congress/senate-bill/4281/text> (not retrievable on August 28, 2026). Supports: MATCH Act (S.4281), bill text with 150-day deadline and threat of sanctions.
[^108]: TrendForce: "[News] U.S. May Push Near-Total ASML China Ban on DUV Sales and Servicing Amid China Domestic Lithography Push". August 25, 2026. <https://www.trendforce.com/news/2026/08/25/news-u-s-may-push-near-total-asml-china-ban-on-duv-sales-and-servicing-amid-chinas-domestic-lithography-push/> (accessed August 28, 2026). Supports: US pushes near-total ban on ASML's China business including servicing; China's share of ASML system revenue 14 % (Q2 2026).
[^109]: NL Times: "U.S. preparing to force Netherlands to ban ASML from selling to China". August 20, 2026. <https://nltimes.nl/2026/08/20/us-preparing-force-netherlands-ban-asml-selling-china> (not retrievable on August 28, 2026). Supports: US preparing to force the Netherlands into an ASML China ban, resistance from The Hague (trade minister Sjoerdsma).
[^110]: Rowan, Linda R.: "Critical Mineral Resources: National Policy and Critical Minerals List". Congressional Research Service. <https://www.congress.gov/crs_external_products/R/PDF/R47982/R47982.13.pdf> (accessed August 28, 2026). Supports: 12 elements (1980s) → more than 60 (NRC time series on Intel data).
[^111]: SIA: "SIA Comments on BIS Section 232 Investigation on Imports of Polysilicon". August 6, 2025. <https://www.semiconductors.org/wp-content/uploads/2025/08/Semiconductor-Industry-Association-SIA-Comments-Polysilicon-Section-232-Investigation.pdf> (not retrievable on August 28, 2026). Supports: semiconductor polysilicon ≥ 11N, \~2 % of volume; Wacker + Hemlock ≈ ¾ of the world market.
[^112]: Bernreuter, Johannes: "Polysilicon Manufacturers: Global Top 10". Bernreuter Research, November 25, 2025. <https://www.bernreuter.com/polysilicon/manufacturers/> (accessed August 28, 2026). Supports: polysilicon manufacturer ranking, China \~95 % of capacity (mostly solar grade), Wacker ranked 8th.
[^113]: IBM: "Copper interconnects". <https://www.ibm.com/history/copper-interconnects> (not retrievable on August 28, 2026). Supports: copper interconnects since 1997.
[^114]: Exiger: "China Announces Export Controls on Five Critical Minerals". February 12, 2025. <https://www.exiger.com/perspectives/critical-minerals-export-controls/> (accessed August 28, 2026). Supports: China's export controls on tungsten and others since 02/04/2025; China controls \~80 % of world supply.
[^115]: IEA: "Global Critical Minerals Outlook 2025". May 2025. <https://www.iea.org/reports/global-critical-minerals-outlook-2025/executive-summary> (accessed August 28, 2026). Supports: cobalt DR Congo \~76 %, China \~90 % of rare-earth refining.
[^116]: Matthews, Jeremy N. A.: "Semiconductor industry switches to hafnium-based transistors". Physics Today, February 1, 2008. <https://physicstoday.aip.org/news/semiconductor-industry-switches-to-hafnium-based-transistors> (accessed August 28, 2026). Supports: semiconductor industry switches to hafnium-based transistors (Intel 45 nm).
[^117]: Baskaran, Gracelin and Schwartz, Meredith: "China’s Antimony Export Restrictions: The Impact on U.S. National Security". CSIS, August 20, 2024. <https://www.csis.org/analysis/chinas-antimony-export-restrictions-impact-us-national-security> (accessed August 28, 2026). Supports: China's antimony export restrictions since 08/2024 and their consequences.
[^118]: Du, Xiaoying and Zhang, Nico: "Antimony 2026: Ample supply, strategic demand to guide geopolitical oversight". Fastmarkets, January 7, 2026. <https://www.fastmarkets.com/insights/antimony-2026-ample-supply-strategic-demand-to-guide-geopolitical-oversight/> (accessed August 28, 2026). Supports: antimony, price from \~14 to \~60 k$/t (peak 07/2025), back to \~23 k$/t in 07/2026.
[^119]: SFA: "The Tin Market". <https://www.sfa-oxford.com/rare-earths-and-minor-metals/minor-metals-and-minerals/tin-market-and-tin-price-drivers/> (accessed August 28, 2026). Supports: tin market and price drivers, solder ≈ 52 % of the market, Myanmar stop, Indonesian restrictions.
[^120]: International Tin Association: "Tin hits nominal all-time-high". January 15, 2026. <https://www.internationaltin.org/tin-hits-nominal-all-time-high/> (accessed August 28, 2026). Supports: tin hits nominal all-time high (\~53,500 $/t).
[^121]: Trading Economics: "Tin - Price - Chart - Historical Data - News". August 28, 2026. <https://tradingeconomics.com/commodity/tin> (accessed August 28, 2026). Supports: tin price \~56,000 $/t.
[^122]: DeCarlo, Samantha and Goodman, Samuel: "Ukraine, Neon, and Semiconductors". USITC, April 2022. <https://www.usitc.gov/publications/332/executive_briefings/ebot_decarlo_goodman_ukraine_neon_and_semiconductors.pdf> (not retrievable on August 28, 2026). Supports: neon, Ukraine supplied \~50 % of world demand and up to \~90 % of US semiconductor demand.
[^123]: Chan, Kelvin: "Iran war cut off helium from Qatar, and shortages will start to bite in a few weeks, threatening chip supply chains that fuel the AI boom". Fortune, March 21, 2026. <https://fortune.com/2026/03/21/iran-war-helium-shortage-qatar-chip-supply-chains-ai-boom/> (accessed August 28, 2026). Supports: helium, US \~43 %, Qatar \~33 %; Hormuz crisis cuts off \~30 % of world supply.
[^124]: Thadani, Akhil and Allen, Gregory C.: "Mapping the Semiconductor Supply Chain: The Critical Role of the Indo-Pacific Region". CSIS, May 30, 2023. <https://csis-website-prod.s3.amazonaws.com/s3fs-public/2023-05/230530_Thadani_MappingSemiconductor_SupplyChain.pdf?VersionId=SK1wKUNf_.qSF3kzMF.aG8dwd.fFTURH> (accessed August 28, 2026). Supports: Japan \~70 % for HF, \~90 % for EUV photoresists; Japan-South Korea export dispute 2019.
[^125]: USGS: "Gallium - Mineral Commodity Summaries 2026". February 2026. <https://pubs.usgs.gov/periodicals/mcs2026/mcs2026-gallium.pdf> (not retrievable on August 28, 2026). Supports: gallium, China \~98–99 % of primary production (official).
[^126]: Blackwood, Matthew and DeFilippo, Catherine: "Germanium and Gallium: U.S. Trade and Chinese Export Controls". USITC, March 2024. <https://www.usitc.gov/publications/332/executive_briefings/ebot_germanium_and_gallium.pdf> (not retrievable on August 28, 2026). Supports: germanium and gallium, China \~60 % for germanium.
[^127]: Wiseman, Paul: "How Will Tight Copper Market Affect Data Center Growth?". Industrial Info, February 20, 2026. <https://www.industrialinfo.com/iirenergy/industry-news/article/how-will-tight-copper-market-affect-data-center-growth--353673> (accessed August 28, 2026). Supports: tight copper market, AI campuses need 3–4× the copper.
[^128]: Boyadjis, Rae: "How AI and data center growth drive copper demand in the US". Fastmarkets, September 16, 2025. <https://www.fastmarkets.com/insights/copper-demand-data-centers-future-trends/> (accessed August 28, 2026). Supports: copper demand from data centers, BNEF deficit \~6 Mt by 2035.
[^129]: S&P Global: "Copper faces 30% supply deficit by 2035, IEA warns at UK summit". December 1, 2025. <https://www.spglobal.com/energy/en/news-research/latest-news/metals/120125-copper-faces-30-supply-deficit-by-2035-iea-warns-at-uk-summit> (not retrievable on August 28, 2026). Supports: IEA warns of a 30 % copper supply deficit by 2035.
[^130]: Treadgold, Tim: "Growing Chinese Control Of Copper A Rare Earths Re-Run". Forbes, August 10, 2026. <https://www.forbes.com/sites/timtreadgold/2026/08/10/growing-chinese-control-of-copper-a-rare-earths-re-run/> (accessed August 28, 2026). Supports: copper all-time high \~15,000 $/t, growing Chinese control.
[^131]: Semiconductor Engineering: "How Semiconductor Fabs Use Water". <https://semiengineering.com/how-semiconductor-fabs-use-water/> (not retrievable on August 28, 2026). Supports: water use of fabs, 20–38 mn liters/day, 30–50 MW, ultrapure water, recycling rates.
[^132]: NPR: "Epic drought in Taiwan pits farmers against high-tech factories for water". April 19, 2023. <https://www.npr.org/sections/goatsandsoda/2023/04/19/1170425349/epic-drought-in-taiwan-pits-farmers-against-high-tech-factories-for-water> (not retrievable on August 28, 2026). Supports: Taiwan's 2021 drought, irrigation shut off for tens of thousands of acres of farmland, chip fabs supplied with priority.
[^133]: Rare Earth Exchanges: "AI’s Hidden Rare Earth Dependency: The Magnets Keeping Data Centers Cool". June 2, 2026. <https://rareearthexchanges.com/news/ais-hidden-rare-earth-dependency-the-magnets-keeping-data-centers-cool/> (accessed August 28, 2026). Supports: rare earths in cooling magnets, China \~85–90 % of processing, \~90 % of magnet production.
[^134]: Sheng, Jenny (Jia) and Hafeez, Sahar J. et al.: "China Suspends Export Controls on Certain Critical Minerals and Related Items". Pillsbury, November 13, 2025. <https://www.pillsburylaw.com/en/news-and-insights/china-suspends-export-controls-certain-critical-minerals-related-items.html> (accessed August 28, 2026). Supports: China suspends the October 2025 control package until 11/27/2026, underlying regime remains active.
[^135]: CNBC: "China suspends ban on exports of gallium, germanium, antimony to U.S.". November 9, 2025. <https://www.cnbc.com/2025/11/09/china-suspends-ban-on-exports-of-gallium-germanium-antimony-to-us.html> (not retrievable on August 28, 2026). Supports: China suspends export ban on gallium, germanium and antimony to the US.
[^136]: Nath, Abhisri and Bean, Jeffrey D.: "China’s Critical Mineral Export Controls: Background & Chokepoints". ORF America, April 22, 2025. <https://orfamerica.org/newresearch/chinas-critical-mineral-export-controls> (accessed August 28, 2026). Supports: China's critical mineral export controls, a pause for recalibration, not reconciliation (analysis).
[^137]: von der Leyen, Ursula: "Opening speech by President von der Leyen at the EU Industry Days 2021". European Commission, February 23, 2021. <https://ec.europa.eu/commission/presscorner/detail/en/speech_21_745> (accessed August 28, 2026). Supports: "98 % of the rare earth elements we need come from a single supplier: China."
[^138]: European Commission: "Critical Raw Materials Act". <https://single-market-economy.ec.europa.eu/sectors/raw-materials/areas-specific-interest/critical-raw-materials/critical-raw-materials-act_en> (accessed August 28, 2026). Supports: Critical Raw Materials Act, in force since 05/23/2024, 2030 targets 10 % extraction / 40 % processing / max. 65 % per third country.
[^139]: Solvay: "Solvay advances European rare earths production". April 8, 2025. <https://www.solvay.com/sites/g/files/srpend616/files/2025-04/20250408%20Inauguration%20La%20Rochelle%20PR%20-%20EN_0.pdf> (not retrievable on August 28, 2026). Supports: magnet rare-earth line in La Rochelle inaugurated, target 30 % of EU demand by 2030.
[^140]: Rahman, Robi and Owen, David: "The training compute of notable AI models has been doubling roughly every six months". Epoch AI, June 19, 2024. <https://epoch.ai/data-insights/compute-trend-post-2010> (accessed August 28, 2026). Supports: training compute of frontier models has doubled roughly every 5–6 months since 2020.
[^141]: Cottier, Ben and Rahman, Robi et al.: "How much does it cost to train frontier AI models?". Epoch AI, June 3, 2024. <https://epoch.ai/blog/how-much-does-it-cost-to-train-frontier-ai-models> (accessed August 28, 2026). Supports: cost of a frontier training run grows \~2.4×/year, >1 bn $ per run by 2027; lion's share hardware, electricity single-digit percent.
[^142]: Sevilla, Jaime and Besiroglu, Tamay et al.: "Can AI scaling continue through 2030?". Epoch AI, August 20, 2024. <https://epoch.ai/blog/can-ai-scaling-continue-through-2030> (accessed August 28, 2026). Supports: scaling to \~2·10²⁹ FLOP possible, 1–5 GW per campus; electricity binds first, then chip manufacturing.
[^143]: Vahdat, Amin and Dean, Jeff: "How much energy does Google’s AI use? We did the math". Google Cloud, August 21, 2025. <https://cloud.google.com/blog/products/infrastructure/measuring-the-environmental-impact-of-ai-inference> (accessed August 28, 2026). Supports: median Gemini text prompt 0.24 Wh / 0.26 ml of water, energy per prompt −33× within 12 months (self-reported).
[^144]: Elsworth, Cooper and Huang, Keguo et al.: "Measuring the environmental impact of delivering AI at Google Scale". arXiv, August 21, 2025. <https://arxiv.org/html/2508.15734v1> (accessed August 28, 2026). Supports: Google's technical paper on the measurement methodology of the inference footprint.
[^145]: Crownhart, Casey: "In a first, Google has released data on how much energy an AI prompt uses". MIT Technology Review, August 21, 2025. <https://www.technologyreview.com/2025/08/21/1122288/google-gemini-ai-energy/> (accessed August 28, 2026). Supports: assessment of Google's figures on energy use per Gemini prompt.
[^146]: McKinsey: "The next big shifts in AI workloads and hyperscaler strategies". December 17, 2025. <https://www.mckinsey.com/industries/technology-media-and-telecommunications/our-insights/the-next-big-shifts-in-ai-workloads-and-hyperscaler-strategies> (not retrievable on August 28, 2026). Supports: AI inference power requirement \~21 GW (2025) → \~93 GW (2030, \~35 % p. a.), training 23 → 62 GW (\~22 % p. a.), projection.
[^147]: DCD: "Training built the campuses. Inference will choose the markets". May 7, 2026. <https://www.datacenterdynamics.com/en/opinions/training-built-the-campuses-inference-will-choose-the-markets/> (not retrievable on August 28, 2026). Supports: "Training built the campuses, inference will choose the markets", inference as latency-bound continuous load.
[^148]: Barabanov, Mykyta: "The AI Infrastructure Bottleneck as a Semiconductor Constraint Rotation". EliteCurrenSea, June 9, 2026. <https://elitecurrensea.com/stocks/the-ai-semiconductor-boom-and-what-could-end-it-bundle-1cprc/> (accessed August 28, 2026). Supports: "The binding link is now the power grid, not silicon", analysis of the AI semiconductor boom.
[^149]: IEA: "Energy and AI". April 2025. <https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai> (accessed August 28, 2026). Supports: data centers \~485 TWh 2025 → \~950 TWh 2030, growing four times faster than any other sector.
[^150]: IEA: "Key Questions on Energy and AI". <https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary> (accessed August 28, 2026). Supports: update "Key Questions on Energy and AI".
[^151]: Pilz, Konstantin F. and Mahmood, Yusuf et al.: "AI’s Power Requirements Under Exponential Growth: Extrapolating AI Data Center Power Demand and Assessing Its Potential Impact on U.S. Competitiveness". RAND. <https://www.rand.org/pubs/research_reports/RRA3572-1.html> (not retrievable on August 28, 2026). Supports: Anthropic estimate, US AI sector needs ≥ 50 GW of additional capacity by 2028.
[^152]: MacroMicro: "Outlook 2026 Series | IV. The AI Power Endgame: The Infrastructure Race from Chips to the Grid". December 17, 2025. <https://en.macromicro.me/blog/outlook-2026-series-iv-the-ai-power-endgame-the-infrastructure-race-from-chips-to-the-grid> (not retrievable on August 28, 2026). Supports: \~11 GW of US capacity stuck at grid connection, multi-year lead times for transformers/switchgear, queues >5 years.
[^153]: CSO Ireland: "Data Centres Metered Electricity Consumption 2024 - Key Findings". June 10, 2025. <https://www.cso.ie/en/releasesandpublications/ep/p-dcmec/datacentresmeteredelectricityconsumption2024/keyfindings/> (accessed August 28, 2026). Supports: data centers = 22 % of Irish electricity consumption in 2024 (official).
[^154]: DCD: "EirGrid says no new applications for data centers in Dublin until 2028 - report". <https://www.datacenterdynamics.com/en/news/eirgrid-says-no-new-applications-for-data-centers-in-dublin-till-2028/> (not retrievable on August 28, 2026). Supports: EirGrid, connection freeze on new data centers in the Dublin area since 2022.
[^155]: Barrett, Eva and Booth, Colm et al.: "Ireland’s Data Centre Connections: Back Online". William Fry, June 24, 2026. <https://www.williamfry.com/knowledge/irelands-data-centre-connections-back-online/> (accessed August 28, 2026). Supports: Ireland's data-center connections back online, conditions: own generation/storage for the full load, 80 % from new Irish renewables within 6 years, feed-back obligation.
[^156]: GE Vernova: "GE Vernova and Crusoe announce major 29-unit aeroderivative gas turbine deal to deliver power to AI data centers". July 22, 2025. <https://www.gevernova.com/news/press-releases/ge-vernova-crusoe-announce-major-29-unit-aeroderivative-gas-turbine-deliver-ai-data-centers> (accessed August 28, 2026). Supports: Crusoe orders 29 aeroderivative gas turbines (\~1 GW) for AI data centers.
[^157]: DCD: "Parker Hannifin to supply equipment for more than 1GW of natural gas turbines to Stargate’s Abilene campus in Texas". <https://www.datacenterdynamics.com/en/news/parker-hannifin-to-supply-more-than-1gw-of-natural-gas-turbines-to-stargates-abilene-campus-in-texas/> (not retrievable on August 28, 2026). Supports: Parker Hannifin to supply >1 GW of gas turbines for the Stargate Abilene campus (on-site generation).
[^158]: Foley, Brynna: "The Queue Before the Queue | GEV’s Backlog Extends to 2031". Enverus, August 4, 2026. <https://www.enverus.com/blog/the-queue-before-the-queue-gevs-backlog-extends-to-2031/> (accessed August 28, 2026). Supports: GE Vernova backlog 116 GW, deliveries through 2031.
[^159]: Martucci, Brian: "GE Vernova expects to end 2025 with an 80-GW gas turbine backlog that stretches into 2029". Utility Dive, December 11, 2025. <https://www.utilitydive.com/news/ge-vernova-gas-turbine-investor/807662/> (accessed August 28, 2026). Supports: GE Vernova, gas-turbine order backlog and lead times (investor update).
[^160]: Energy Connects: "Siemens Energy’s gas turbine boom brives highest-ever order backlog of €136bn". August 11, 2025. <https://www.energyconnects.com/news/oil/2025/august/siemens-energy-s-gas-turbine-boom-brives-highest-ever-order-backlog-of-136bn/> (accessed August 28, 2026). Supports: Siemens Energy record order backlog 136 bn €, gas sold out through 2028, \~4 years' waiting time (CEO Bruch, self-reported).
[^161]: Penrod, Emma: "Siemens Energy’s gas turbine backlog nears 70 GW as company expands manufacturing". Utility Dive, August 10, 2026. <https://www.utilitydive.com/news/siemens-gas-turbine-backlog-nears-70-gw-as-company-expands-manufacturing/827390/> (accessed August 28, 2026). Supports: Siemens Energy gas-turbine backlog \~69 GW, lead times 3+ years.
[^162]: Holt, Marin: "Siemens Energy’s Grid Order Backlog Hits a Record 51 Billion Euros on Transformer Demand". mgrid, August 9, 2026. <https://mgrid.org/2026/08/09/siemens-energys-grid-order-backlog-hits-a-record-51-billion-euros-on-transformer-demand/> (accessed August 28, 2026). Supports: Siemens Energy grid order backlog a record 51 bn € on transformer demand; individual transformer orders >5 years.
[^163]: Penrod, Emma: "Mitsubishi’s large-frame gas turbine backlog reaches 35 GW". Utility Dive, August 13, 2026. <https://www.utilitydive.com/news/mitsubishi-gas-turbine-backlog-earnings/827761/> (accessed August 28, 2026). Supports: Mitsubishi Heavy Industries, large-frame turbine backlog 35 GW, new orders delivering 2028–2030, "selective in the projects we contract".
[^164]: Robb, Drew: "Data Centers Look to Old Airplane Engines for Power". IEEE Spectrum, October 20, 2025. <https://spectrum.ieee.org/ai-data-centers> (accessed August 28, 2026). Supports: AI data centers, heavy-duty turbines \~3–5 years' lead time, aeroderivatives now also 3–5 years' ordering lead time (industry figures).
[^165]: NucNet: "Constellation Secures $1 Billion Federal Loan For Three Mile Island Restart". November 3, 2025. <https://www.nucnet.org/news/constellation-secures-usd1-billion-federal-loann-for-three-mile-island-restart-11-3-2025> (not retrievable on August 28, 2026). Supports: Constellation secures 1 bn $ DOE loan for the Three Mile Island restart ("Crane Clean Energy Center"), target brought forward to 2027.
[^166]: Crosley, Blake: "Nuclear power for AI: inside the data center energy deals". Introl, January 8, 2026. <https://introl.com/blog/nuclear-power-ai-data-centers-microsoft-google-amazon-2025> (accessed August 28, 2026). Supports: nuclear deals AWS-Talen (Susquehanna, up to 1.92 GW), Microsoft-Constellation, Google-Kairos.
[^167]: Patel, Sonal C.: "The SMR Gamble: Betting on Nuclear to Fuel the Data Center Boom". POWER Magazine, March 3, 2025. <https://www.powermag.com/the-smr-gamble-betting-on-nuclear-to-fuel-the-data-center-boom/> (accessed August 28, 2026). Supports: SMR market \~6.9 bn $ in 2025, nuclear for data centers only from 2027+.
[^168]: insideHPC: "‘Plug, Baby, Plug’: France to Use Nuclear Power to Expand AI Computing Capacity". February 2025. <https://insidehpc.com/2025/02/plug-baby-plug-france-to-use-nuclear-power-to-expand-its-ai-computing-capacity/> (not retrievable on August 28, 2026). Supports: Macron "plug, baby, plug", France markets nuclear power to AI data centers.
[^169]: pv magazine USA: "Redwood, Crusoe deploy second-life batteries at AI data center for 63 MWh storage". June 30, 2025. <https://pv-magazine-usa.com/2025/06/30/redwood-crusoe-deploy-second-life-batteries-at-ai-data-center-for-63-mwh-storage/> (not retrievable on August 28, 2026). Supports: Redwood/Crusoe, microgrid of 12 MW solar and 63 MWh second-life EV batteries in Nevada, largest second-life deployment worldwide.
[^170]: Rayner, Tristan: "Second-life EV batteries get tick from US data center operator". ESS News, March 25, 2026. <https://www.ess-news.com/2026/03/25/second-life-ev-batteries-get-tick-from-us-data-center-operator/> (accessed August 28, 2026). Supports: 99.2 % availability over seven months in the Redwood/Crusoe microgrid (self-reported).
[^171]: Crusoe: "Crusoe and Redwood Materials Expand Strategic Partnership, Scaling to 7x the Original AI Infrastructure Density". March 24, 2026. <https://www.crusoe.ai/resources/newsroom/crusoe-and-redwood-materials-expand-strategic-partnership-scaling-to-7x-the-original-ai-infrastructure-density> (accessed August 28, 2026). Supports: build-out from 4 to 24 modules (\~7× compute) with Redwood (self-reported).
[^172]: O’Dea, Blathnaid: "Masdar, EWEC announce 5 GW/19 GWh solar-plus-storage project in Abu Dhabi". pv magazine, January 14, 2025. <https://www.pv-magazine.com/2025/01/14/masdar-ewec-announce-5-gw-19-gwh-solar-plus-storage-project-in-abu-dhabi/> (accessed August 28, 2026). Supports: Masdar/EWEC announce 5 GW PV + 19 GWh storage for 1 GW baseload in Abu Dhabi.
[^173]: Shaw, Vincent: "Sungrow wins 7.5 GWh Masdar order for Abu Dhabi 24/7 solar-storage project". ESS News, May 22, 2026. <https://www.ess-news.com/2026/05/22/sungrow-wins-7-5-gwh-masdar-order-for-abu-dhabi-24-7-solar-storage-project/> (accessed August 28, 2026). Supports: Sungrow to supply 7.5 GWh of storage for Masdar's 24/7 solar project.
[^174]: Lambert, Fred: "BYD wins 11.3 GWh battery deal for world’s largest solar+storage". Electrek, July 10, 2026. <https://electrek.co/2026/07/10/byd-masdar-11-gwh-rtc-storage-abu-dhabi/> (accessed August 28, 2026). Supports: BYD to supply 11.3 GWh of storage for Masdar; 6.1 bn $ investment, financial close.
[^175]: Norman, Will: "Masdar breaks ground on ‘world’s largest’ 5.2GW/19GWh solar-plus-storage project". PV Tech, October 24, 2025. <https://www.pv-tech.org/masdar-breaks-ground-on-worlds-largest-5-2gw-19gwh-solar-plus-storage-project/> (accessed August 28, 2026). Supports: Masdar breaks ground on the world's largest 5.2 GW / 19 GWh solar-plus-storage project, operation from 2027.
[^176]: pv magazine USA: "U.S. transformer market faces severe supply constraints as lead times extend to four years". May 11, 2026. <https://pv-magazine-usa.com/2026/05/11/u-s-transformer-market-faces-severe-supply-constraints-as-lead-times-extend-to-four-years/> (not retrievable on August 28, 2026). Supports: US transformer market, lead times extend to four years.
[^177]: EEPower: [Title not determinable]. <https://eepower.com/tech-insights/transformer-supply-chain-woes-persist-as-energy-demand-grows/> (not retrievable on August 28, 2026). Supports: transformer supply chain, US imports \~80 % of large power transformers (trade press).
[^178]: Beaubouef, Bruce: "Shortage of submarine power cables poses threat to offshore wind market". Offshore Magazine, November 4, 2024. <https://www.offshore-mag.com/renewable-energy/article/55240474/shortage-of-submarine-power-cables-poses-threat-to-offshore-wind-market> (accessed August 28, 2026). Supports: shortage of high-voltage submarine cables, NKT/Prysmian/Nexans \~75 % market share, backlogs "12 years plus".
[^179]: SGB-SMIT: "SGB-SMIT Group - Manufacturer of transformers in europe". <https://www.sgb-smit.com/> (accessed August 28, 2026). Supports: independent transformer builder, Regensburg.
[^180]: Maschinenfabrik Reinhausen: "THE POWER BEHIND POWER | Energy technology for the future". <https://www.reinhausen.com/> (accessed August 28, 2026). Supports: tap changers, \~50 % of the world market (manufacturer figure).
[^181]: Inside Climate News: [Title not determinable]. July 17, 2025. <https://insideclimatenews.org/news/17072025/elon-musk-xai-data-center-gas-turbines-memphis/> (not retrievable on August 28, 2026). Supports: xAI Memphis, up to 35 mobile gas turbines, 15 permitted, Clean Air Act lawsuit.
[^182]: Ireland, Illan: "Mississippi Permit Board Grants xAI’s Request for 41 Southaven Gas Turbines to Power Memphis Data Center". Mississippi Free Press, March 10, 2026. <https://www.mississippifreepress.org/mississippi-permit-board-grants-xais-request-for-41-southaven-gas-turbines-to-power-memphis-data-center/> (accessed August 28, 2026). Supports: permit board grants xAI 41 permanent gas turbines in Southaven for Colossus 2 (1.2 GW).
[^183]: GE Vernova: "Homer City Redevelopment and Kiewit announce country’s largest natural gas-powered data center campus to support AI and HPC demand". April 2, 2025. <https://www.gevernova.com/news/press-releases/homer-city-redevelopment-kiewit-announce-country-largest-natural-gas-powered-data-center-support-ai-hpc-demand> (accessed August 28, 2026). Supports: Homer City, former coal plant becomes the largest gas-powered data-center campus in the US (up to 4.5 GW, 7× GE 7HA).
[^184]: Blake, Mathias and Hsu, Martin et al.: "NVIDIA 800 VDC Architecture Will Power the Next Generation of AI Factories". Nvidia, May 20, 2025. <https://developer.nvidia.com/blog/nvidia-800-v-hvdc-architecture-will-power-the-next-generation-of-ai-factories/> (accessed August 28, 2026). Supports: 800 V HVDC architecture, in series from 2027 (Kyber), \~5 % efficiency gain (manufacturer figure).
[^185]: Hitachi Energy: "Hitachi Energy to invest additional $4.5 billion by 2027 to accelerate the clean energy transition". June 7, 2024. <https://www.hitachienergy.com/news-and-events/press-releases/2024/06/hitachi-energy-to-invest-additional-4-5-billion-by-2027-to-accelerate-the-clean-energy-transition> (accessed August 28, 2026). Supports: additional 4.5 bn $ investment by 2027 (expansion program \~6 bn $ in total).
[^186]: Walton, Robert: "Hitachi Energy commits $250M to address transformer shortage". Utility Dive, March 10, 2025. <https://www.utilitydive.com/news/hitachi-energy-commits-250-million-transformer-shortage/742010/> (accessed August 28, 2026). Supports: Hitachi Energy commits a further 250 mn $ against the transformer shortage.
[^187]: Patel, Sonal C.: "Transformers in 2026: Shortage, Scramble, or Self-Inflicted Crisis?". POWER Magazine, January 2, 2026. <https://www.powermag.com/transformers-in-2026-shortage-scramble-or-self-inflicted-crisis/> (accessed August 28, 2026). Supports: transformers, lead time China \~12 months vs. US 128 weeks / EU 48–60 months; China >50 % of world production, TBEA No. 1.
[^188]: Karaahmetovic, Senad: "Chinese transformer exports to US data centers surge, Morgan Stanley reports". Investing.com, November 12, 2025. <https://www.investing.com/news/stock-market-news/chinese-transformer-exports-to-us-data-centers-surge-morgan-stanley-reports-93CH-4351391> (accessed August 28, 2026). Supports: Chinese transformer exports 2025 +36 %, US data centers the largest buyer.
[^189]: Global Energy Monitor: "Ultra-High-Voltage (UHV) Power Transmission System in China". August 27, 2026. <https://www.gem.wiki/Ultra-High-Voltage_(UHV)_Power_Transmission_System_in_China> (accessed August 28, 2026). Supports: China's UHV grid, 19 AC + 20 DC projects, >40,000 km.
[^190]: Fairley, Peter: "China’s State Grid Corp Crushes Power Transmission Records". IEEE Spectrum, January 10, 2019. <https://spectrum.ieee.org/chinas-state-grid-corp-crushes-power-transmission-records> (accessed August 28, 2026). Supports: State Grid, Changji-Guquan ±1,100 kV, 12 GW over 3,324 km.
[^191]: Martos, Jenny: "Leading three manufacturers providing two-thirds of turbines for gas-fired power plants under construction". Global Energy Monitor, August 1, 2024. <https://globalenergymonitor.org/research/leading-three-manufacturers-providing-two-thirds-turbines-gas-fired-power-plants-under> (accessed August 28, 2026). Supports: three manufacturers provide two-thirds of gas turbines under construction, in China GE/Harbin \~39 %, Dongfang/Mitsubishi \~25 % (licensed manufacture).
[^192]: China Daily: "First self-developed 300 MW F-class heavy-duty gas turbine completes ignition test". October 8, 2024. <https://www.chinadaily.com.cn/a/202410/08/WS6704e993a310f1265a1c6797.html> (accessed August 28, 2026). Supports: Shanghai Electric, gas-turbine manufacturing under Ansaldo GT36 license.
[^193]: SASAC: [Title not determinable]. January 8, 2026. <http://en.sasac.gov.cn/2026/01/08/c_20296.htm> (not retrievable on August 28, 2026). Supports: first indigenous 300 MW F-class gas turbine in commercial operation since 12/29/2025 (Huadian Wangting, unit 5), official.
[^194]: Ember: [Title not determinable]. <https://ember-energy.org/countries-and-regions/china/> (not retrievable on August 28, 2026). Supports: gas \~2–3 % of China's electricity mix.
[^195]: Bloomberg: "China’s Record Renewables Buildout Is Wasting Power as Grid Lags". August 5, 2025. <https://www.bloomberg.com/news/articles/2025-08-05/china-s-record-renewables-buildout-is-wasting-power-as-grid-lags> (not retrievable on August 28, 2026). Supports: China's record renewables build-out is wasting power as the grid lags.
[^196]: Myllyvirta, Lauri: "Analysis: China’s CO2 climbs 2% in early 2026 due to ‘wasted’ wind and solar". Carbon Brief, June 4, 2026. <https://www.carbonbrief.org/analysis-chinas-co2-climbs-2-in-early-2026-due-to-wasted-wind-and-solar> (accessed August 28, 2026). Supports: curtailment H1 2026, solar 8.6 %, wind 9.1 % (NEA, official); including unreported curtailment \~26 % (CREA/GEM estimate).
[^197]: Google Research: "Meet Project Suncatcher, a research moonshot to scale machine learning compute in space". November 4, 2025. <https://blog.google/innovation-and-ai/technology/research/google-project-suncatcher/> (accessed August 28, 2026). Supports: Project Suncatcher, up to \~8× solar yield in sun-synchronous orbit, radiator cooling, prototype mission with Planet Labs \~early 2027; economics hinge on launch costs (research figure).
[^198]: DCD: [Title not determinable]. November 2025. <https://www.datacenterdynamics.com/en/news/starcloud-1-satellite-reaches-space-with-nvidia-h100-gpu-now-operating-in-orbit/> (not retrievable on August 28, 2026). Supports: Starcloud-1 puts the first Nvidia H100 into orbit.
[^199]: CNBC: [Title not determinable]. December 10, 2025. <https://www.cnbc.com/2025/12/10/nvidia-backed-starcloud-trains-first-ai-model-in-space-orbital-data-centers.html> (not retrievable on August 28, 2026). Supports: Starcloud trains the first AI model in space.
[^200]: Fernholz, Tim: "Starcloud raises $250 million for orbital data centers as launch options dry up". TechCrunch, August 21, 2026. <https://techcrunch.com/2026/08/21/starcloud-raises-200-million-for-orbital-data-centers-as-launch-options-dry-up/> (accessed August 28, 2026). Supports: Starcloud funding round (+250 mn $ at 2.3 bn $, Nvidia participating); launch capacity drying up, Falcon 9 program ends in 2028.
[^201]: GeekWire: [Title not determinable]. August 2026. <https://www.geekwire.com/2026/starcloud-250m-data-center-satellite-network-nvidia/> (not retrievable on August 28, 2026). Supports: Starcloud plans a space-grade "Vera Rubin Space-1" GPU from late 2028, Starcloud-2 in 2027 (self-reported).
[^202]: Nvidia: "NVIDIA Launches Space Computing, Rocketing AI Into Orbit". March 16, 2026. <https://nvidianews.nvidia.com/news/space-computing> (accessed August 28, 2026). Supports: Space-1 Vera Rubin module for orbital computing (2027).
[^203]: Subramanian, Pras: "SpaceX reveals its first orbital data center, ‘much simpler than a Starlink satellite,’ Musk says". Yahoo Finance, June 9, 2026. <https://finance.yahoo.com/sectors/technology/article/spacex-reveals-its-first-orbital-data-center-much-simpler-than-a-starlink-satellite-musk-says-141110185.html> (accessed August 28, 2026). Supports: SpaceX reveals its first orbital data-center satellite "AI1" (Musk, self-reported).
[^204]: Crypto Briefing: "SpaceX reveals AI1 orbital data center design for satellite network". July 15, 2026. <https://cryptobriefing.com/spacex-ai1-orbital-data-center-satellite-network/> (accessed August 28, 2026). Supports: SpaceX constellation "Starmind", FCC filing for up to 1 mn satellites.
[^205]: Bloomberg: [Title not determinable]. August 24, 2026. <https://www.bloomberg.com/news/articles/2026-08-24/spacex-s-musk-sees-orbital-data-center-launch-near-end-of-2027> (not retrievable on August 28, 2026). Supports: Musk, first orbital data-center launches at the end of 2027 with Nvidia Rubin hardware, scaling in 2028 (self-reported).
[^206]: Fernholz, Tim: "Jeff Bezos’ Blue Origin enters the space data center game". TechCrunch, March 20, 2026. <https://techcrunch.com/2026/03/20/jeff-bezos-blue-origin-enters-the-space-data-center-game/> (accessed August 28, 2026). Supports: Blue Origin files "Project Sunrise" (>50,000 satellites) with the FCC.
[^207]: Thales Alenia Space: "Thales Alenia Space reveals results of ASCEND feasibility study on space data centers". June 27, 2024. <https://www.thalesaleniaspace.com/en/press-releases/thales-alenia-space-reveals-results-ascend-feasibility-study-space-data-centers-0> (accessed August 28, 2026). Supports: results of the EU feasibility study ASCEND on data centers in orbit.
[^208]: Pamir LLC: "Space becomes the new data center frontier as China launches its first 12 satellites for interstellar AI processing". June 10, 2025. <https://pamirllc.com/blog/space-becomes-the-new-data-center-frontier-as-china-launches-its-first-12-satellites-for-interstellar-ai-processing> (accessed August 28, 2026). Supports: Zhejiang Lab launches the first 12 computing satellites of the "Three-Body Computing Constellation" (05/14/2025), target 100 by 2027.
[^209]: SCIO: [Title not determinable]. April 27, 2026. <http://english.scio.gov.cn/chinavoices/2026-04/27/content_118464689.html> (not retrievable on August 28, 2026). Supports: China's space computing constellations, planning status April 2026 (official).
[^210]: Xinhua: "China Focus: From ground to orbit: China eyes computing in space". April 25, 2026. <https://english.news.cn/20260425/b4368923827e4988af1bd030dd01d9cc/c.html> (accessed August 28, 2026). Supports: ADAspace "Star Compute" constellation, 2,800 computing satellites planned (2,400 inference, 400 training).
[^211]: SpaceSGP: "The Orbital Compute Contest: US-China Space Data Centers and Three Windows for Asia-Pacific". April 28, 2026. <https://www.spacesgp.com/en/insights/orbital-compute-us-china-apac-2026> (accessed August 28, 2026). Supports: orbital compute US/China/APAC, ADAspace runs Qwen3 in orbit (11/2025), orbital parameters of the constellations.
[^212]: ChinaTechNews: "China launches first of 1,000 space-computing satellites". July 18, 2026. <https://www.chinatechnews.com/2026/07/18/125844-china-launches-first-of-1000-space-computing-satellites> (accessed August 28, 2026). Supports: launch of the first satellites of a further constellation targeting 1,000 computing satellites.
[^213]: SpaceDaily: "In 2026, China put more than 100 space-computing organizations under a new Beijing committee, tying 12 AI satellites and a 1,000-POPS orbital supercomputer plan to the Five-Year Plan and the still-unsolved problem of moving data through orbit". June 9, 2026. <https://spacedaily.com/sd-in-2026-china-put-more-than-100-space-computing-organizations-under-a-new-beijing-committee-tying-12-ai-satellites-and-a-1000-pops-orbital-supercomputer-plan-to-the-five-year-plan-and-the-still/> (accessed August 28, 2026). Supports: Beijing puts >100 space-computing organizations under a new committee; 1,000-POPS orbital supercomputer in the five-year plan.